0% found this document useful (0 votes)

33 views15 pages

Nudging Towards Responsible Recommendations A Graph-Based Approach To Mitigate Belief Filter Bubbles

Uploaded by

2467953405lzy

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

33 views15 pages

Nudging Towards Responsible Recommendations A Graph-Based Approach To Mitigate Belief Filter Bubbles

Uploaded by

2467953405lzy

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 15

This article has been accepted for publication in IEEE Transactions on Artificial Intelligence.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TAI.2024.3373392

JOURNAL OF IEEE TRANSACTIONS ON ARTIFICIAL INTELLIGENCE, VOL. 00, NO. 0, NOV. 2023 1

Nudging Towards Responsible Recommendations: a

Graph-Based Approach to Mitigate Belief Filter
Bubbles
Mengyan Wang, Yuxuan Hu, Shiqing Wu, Weihua Li, Quan Bai, Zihan Yuan, and Chenting Jiang

Abstract—Personalized recommendation systems homogenize I. I NTRODUCTION

user preferences, causing an extreme belief imbalance and
aggravating user bias. This phenomenon is known as the fil-
ter bubble. This paper presents the Responsible Graph-based
Recommendation (RGRec) system, designed to alleviate the filter
R ECOMMENDATION systems play an important role
in shaping users’ access to diverse information on the
Internet [1]. However, most of the existing preference-based
bubble effect in personalized recommendation systems. Acting as recommendation systems continuously suggest items that are
an intermediate agency between users and existing preference- similar to users’ previous experiences, leading them to a
based recommendation systems, RGRec is composed of three
collaborative modules: the Multi-faceted Reasoning-based Filter
homogeneous information environment. This is known as the
Bubbles Detection module (FBDetect), the Belief Nudging mod- “filter bubble” phenomenon [2]. Prolonged exposure to such
ule, and the Generative Artificial Intelligence-based Recommen- filter bubbles can result in the development of extreme and im-
dation Strategy Generation module (RecomGen). The FBDetect balanced beliefs, hindering the formation of a comprehensive
module identifies users with extreme belief imbalances based on understanding and amplifying ideological biases [3]. In the
their belief networks, which are represented as heterogeneous
graphs. These graphs are then utilized in the Belief Nudging
context of responsible artificial intelligence (AI), it becomes
module, where a nudging strategy is employed to adapt prompts crucial to develop AI recommendation systems that provide
for the RecomGen module. Ultimately, the RecomGen module a diversity of content and viewpoints, rather than reinforcing
generates contextually rich items for recommendations. Exper- existing user preferences, to mitigate filter bubbles [4].
imental results demonstrate that RGRec can promote diverse Existing research works on mitigating the filter bubble in
content exploration based on user feedback and progressively
stimulate interest in topics users initially showed less interest in,
recommendation systems can be grouped into two main strate-
encouraging individual exploration. gies: algorithm-focused and human-focused approaches [5].
Algorithm-focused strategies advocate for promoting content
Impact Statement—This paper introduces an innovative ap-
proach to personalized recommendations, termed Responsible diversity at both the in-processing and post-processing stages
Graph-based Recommendation (RGRec). RGRec addresses the [5]. These methods leverage diverse techniques, including
filter bubble issue associated with recommendation systems by explanation-based diversity recommendation [6], community-
integrating Generative Artificial Intelligence (GAI). Notably, aware models [7], category-based diversification algorithms
RGRec excels in producing contextually rich recommendations [8], the Diversified GNN-based Recommendation system
through graph-based nudging strategies. By considering user
preferences, it actively encourages diverse content exploration. (DGRec) [9], and graph-based user-item interaction meth-
This responsible methodology aligns with principles of liber- ods [10]. Notably, graph-based approaches, relying on user
tarian paternalism and transparency, distinguishing itself from preferences and category diversity insights, aim to enhance
conventional recommendation systems. The proposed approach recommendation quality [11]. However, these strategies, par-
represents a noteworthy advancement in responsible AI and ticularly those driven by algorithms, may unintentionally ig-
recommendation systems, challenging established norms by dis-
mantling filter bubbles, fostering information diversity, and nore the essential role of human decision-making processes.
upholding user autonomy. In contrast, human-focused strategies emphasize individuals
Index Terms—Belief harmony, Filter bubble, Nudge theory, [5]. Techniques such as nudging-based recommendations [12],
Responsible recommendation systems [13] indirectly influence user decisions and behaviors. De-
spite focusing on users, these models often encounter chal-
Mengyan Wang, Auckland University of Technology, Auckland, New lenges in effectively broadening users’ interests beyond their
Zealand (e-mail: [email protected]).
Yuxuan Hu, University of Tasmania, Hobart, Australia (yux-
preferred topics. Unlike existing mitigation approaches, this
[email protected]). paper integrates the strengths of both algorithmic and human-
Shiqing Wu, University of Technology Sydney, Sydney, Australia focused strategies and presents the Responsible Graph-based
([email protected]).
Corresponding author. Weihua Li, Auckland University of Technology,
Recommendation framework, namely RGRec, to mitigate filter
Auckland, New Zealand (e-mail: [email protected]). bubbles by mildly moderating users’ extreme beliefs, thus
Quan Bai, University of Tasmania, Hobart, Australia exposing them to a more diverse scope of information.
([email protected]).
Zihan Yuan, University of Tasmania, Hobart, Australia
RGRec is a graph-based approach that serves as an interme-
([email protected]). diary between recommendation systems and users. Its primary
Chenting Jiang, University of Tasmania, Hobart, Australia (chent- objective is to effectively address the filter bubbles issues by
[email protected]).
bridging the gap between user preferences and the delivery

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on April 08,2024 at 05:23:59 UTC from IEEE Xplore. Restrictions apply.
© 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Artificial Intelligence. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TAI.2024.3373392

2 JOURNAL OF IEEE TRANSACTIONS ON ARTIFICIAL INTELLIGENCE, VOL. 00, NO. 0, NOV. 2023

of diverse content recommendations. The system comprises offers a comprehensive review of related literature, delving
three key modules: the Multi-faceted Reasoning-based Filter into recommendation systems, user belief bias, filter bubbles,
Bubbles Detection module (FBDetect), the Belief Nudging and nudging techniques. Section III elucidates essential defi-
module, and the Generative Artificial Intelligence-based Rec- nitions, notations, and concepts integral to our discourse. The
ommendation Strategy Generation module (RecomGen). In methodology and framework underpinning our research are
FBDetect, a user’s belief is represented as a heterogeneous expounded upon in Section IV. Section V is dedicated to
graph [14] known as a belief network. FBDetect identifies the elucidation of our experimental setup, with subsequent
users affected by filter bubbles by evaluating the balance be- presentation and analysis of results. In Section VI, the findings
tween a user’s belief toward a specific topic of information and are examined and discussed. Section VII concludes the paper
the recommendations received from the system. If this balance and outlines potential directions for future research.
is significantly skewed, the user is flagged as being impacted
by filter bubbles. The Belief Nudging module collaborates II. R ELATED W ORKS
with users’ belief networks to explore paths between topics Personalized recommendation systems have been criticized
that users favor and those they display less interest. These as inadvertently creating filter bubbles [15], which constrain
explored paths serve as prompts for RecomGen to generate users’ exposure to various perspectives and information, thus
items for a nudging recommendation strategy. This strategy potentially leading to belief biases and societal fragmentation
aims to gently introduce users to content they may have shown [8]. To mitigate this concern, many researchers and practi-
less preference for, fostering a more balanced exposure to tioners have focused on dismantling filter bubbles, fostering
diverse content. The collaboration among these three modules diversity and democracy in recommendation systems, and
is iterative, continuously optimizing and adjusting to mitigate facilitating users’ belief harmony.
filter bubbles. In this section, we review the relevant research works,
The ultimate goal of this research is to gradually introduce deliberate on the filter bubble issue, explore the diversification
users to a more diverse range of content, fostering belief of recommendation systems, and examine the prior research
harmony and enhancing the overall user experience. RGRec’s on nudge recommendations. Additionally, we will highlight
innovative approach addresses the challenges of filter bubbles the contributions of this study.
by proactively diversifying recommendations and promoting a
more nuanced interaction between users and content.
A. Filter Bubbles
• Firstly, we introduce a novel responsible approach, i.e.,
RGRec, designed to address the moderation of users’ 1) Preference-based Recommendation Systems: Conven-
extreme beliefs and the mitigation of filter bubbles re- tional recommendation systems prioritize the generalization of
sulting from conventional recommendation approaches. user preference, implying that these systems often recommend
To the best of our knowledge, RGRec stands out as one items to users based on their specific preferences and behaviors
of the pioneering responsible recommendation methods [16]. Techniques such as Collaborative Filtering (CF) [17],
explicitly focused on alleviating filter bubbles. Content-Based filtering (CB) [18], rule-based methods [19], or
• Secondly, we present the Multi-faceted Reasoning-based hybrid models [20] are commonly employed to analyze users’
Filter Bubbles Detection module (FBDetect), a pivotal preferences and past behaviors. The recommendation system
component within RGRec. FBDetect identifies users af- then suggests content that aligns closely with user preferences
fected by filter bubbles and scrutinizes recommendation to enhance user satisfaction and engagement. However, this
systems relying solely on user preferences. Our approach approach based on user preference may exacerbate the filter
employs diverse methodologies to comprehensively anal- bubble issue, leading to ideological isolation and user bias.
yse filter bubbles, examining their existence and effects For example, Bryant et al. demonstrate that the YouTube
from various perspectives. algorithm, representative of a preference recommendation al-
• Thirdly, we leverage the efficacy of nudging techniques gorithm, exhibits a marked bias towards right-leaning political
to guide users in broadening their interests and promot- videos, including those espousing racist views propagated by
ing belief harmony. Our nudging process aligns with the alt-right community [21]. Thus, it is important to address
principles of libertarian paternalism, transparency, and the limitations of current preference recommendations, boost
democracy, thereby enhancing users’ understanding of the diversity of suggestions, and harmonize users’ beliefs.
recommendations. 2) Mitigating Filter Bubble Effects: Filter bubbles empha-
• Finally, we present the Generative Artificial Intelligence- size the constraints of preference recommendation algorithms
based Recommendation Strategy Generation module (Re- [22]. Dahlgren introduced the term “internet filters” to rep-
comGen) for crafting recommendation strategies aimed at resent the phenomenon of filter bubbles, which can have
mitigating filter bubbles. This method leverages advanced various negative effects on users, including a narrowed focus
graph-based techniques to learn and analyze user beliefs, on personal interests, substantial reinforcement of confirmation
systematically exploring potential paths to alleviate users’ bias, reduced curiosity, decreased exposure to diverse ideas
extreme beliefs by introducing a more extensive range of and people, compromised understanding of the world, and a
information and enhancing content diversity. skewed perception of reality [23].
The rest of this paper is organized as follows. Section II Addressing the negative effects of filter bubbles proves to
be challenging, especially when considering the notable aspect

WANG et al.: NUDGING TOWARDS RESPONSIBLE RECOMMENDATIONS 3

of algorithmic bias. Chen et al. argue that the emergence the focus of SNA on alleviating echo-chamber effects rather
of recommendation algorithm bias amplifies the experimental than filter bubbles, our research concentrates on investigating
nature of user behavior data as opposed to observational [24]. selective belief bias algorithms based on ISDA and UIPA.
Additionally, Dahlgren examines the recommendation algo- ISDA includes various detection metrics such as topology
rithm bias and broadens the concept of bias into two facets, metrics and homophily metrics [8]. Likewise, UIPA includes
one originating from the recommendation algorithm and the several established detection metrics, including the coverage
other from users’ behaviors [23]. Aside from algorithmic bias, algorithm and the Majority Category Domination (MCD)
another challenge in mitigating filter bubbles lies in their algorithm [29]. Drawing inspiration from these metrics, we
elusive nature [8]. Users often find themselves unaware of the propose the FBDetect model for dual verification of the au-
filter bubble effect, which creates a homogenized view of the thenticity of the filter bubble phenomenon, having the concept
world. Specifically, they may not realize that their perspective of “Entropy” [31] included to substantiate the existence of
differs from others in similar circumstances. filter bubbles.
The growing influence of filter bubbles has raised increased
concerns among researchers. A well-crafted recommendation B. Nudge Techniques and Responsible Recommendations
system usually offers high accuracy while promoting diversity;
A nudge is a non-coercive intervention designed to influence
systems oriented solely towards accuracy may inevitably lead
behavior by modifying the context in which choices are made
to filter bubble effects [9]. Contemporary research proposes
[32]. Such an intervention is usually transparent, optional, and
several strategies for breaking filter bubbles by enhancing the
responsible, enabling individuals to understand their choice
diversity of recommendations. The research addressing filter
consequences better and boost the likelihood of beneficial
bubbles with graphs is reviewed as follows.
decision-making [33]. The core idea behind a nudge is to
Research in the scope of the graph. While the above-
exploit individuals’ beliefs and behavioral biases through
mentioned research has significantly contributed to alleviating
various design strategies, such as providing incentives [34],
filter bubbles and enhancing recommendation diversity, many
[35] and utilizing social influence [36], [37], directing them
researchers also advocate for the critical role of graph-based
towards more favorable outcomes without constraining their
recommendation algorithms. These algorithms mitigate data
freedom of choice [32], [38].
sparsity and cold start issues and add an essential inter-
Recent research in recommendation systems has begun to
pretability factor to recommendation systems [11]. Yang et
explore the role of nudges. However, the majority introduces
al. introduce the Diversified GNN-based Recommendation
nudging recommendations from an AI-deprived perspective,
System (DGRec), a graph-based recommendation system built
implying a substantial absence or lack of AI technology in
on GNN, augmenting the diversity of recommended lists by
their research context. For example, Jesse et al. consolidate
improving the embedding generation process [9]. Tang et
87 nudging mechanisms at this AI-deprived level, including
al. propose a temporal graph-based method to learn user
alterations in font size, the reputation of the messenger, and the
evolving preferences in dynamic recommendation scenarios
visibility of information [12]. Joachim et al. propose a platform
[25]. Additionally, Li et al. adopt a graph-based methodology
empowered by AI designed to nudge, influence, and guide the
by constructing a user-item interaction graph for data analysis
behavior of individuals with diabetes [13]. Furthermore, Sitar
to examine the existence of a centralized recommendation
et al. propose an automated recommendation system. This
phenomenon [10].
system integrates managers’ priorities and user feedback and
In contrast, our model surpasses traditional methods by
utilizes graph structures to organize items based on descending
generating more diverse items instead of marginally varied
order of priority, known as nudge concepts [39].
ones, based on user preferences for diversity. We prioritize
The recommendation systems mentioned previously have
incrementally stimulating users’ interest in items they may
revealed the importance of establishing a responsible, graph-
initially disregard without altering existing recommendation
based nudging recommendation system. However, existing
system algorithms. The proposed novel approach aims to
models are developed solely on user preferences, neglecting
counteract the filter bubble effect by considering user interest
the influence of filter bubbles. Unlike the existing preference-
and disinterest beliefs, i.e., an aspect that has received minimal
based approaches, this research aims to gently present more
attention from researchers.
potential interests to users whom they may have yet to be
Detection of Belief Bias. Belief bias in reasoning refers
genuinely interested in initially. Guiding user perceptions from
to individuals’ tendency to favor conclusions that align with
one end of the graph (items users are highly interested in) to
their pre-existing beliefs [26]. This phenomenon is connected
the other end (items users are less interested in), moderating
with the formation of online filter bubbles, in which users
user extreme beliefs, reducing user bias, and breaking the filter
tend to accept information that confirms their viewpoints and
bubble effect, thereby allowing users to access a more diverse
interests while rejecting alternative perspectives that challenge
range of information.
their beliefs [27], [28].
Existing methods proposed for belief bias detection include
Information Source Diversity Analysis (ISDA) [8], User Inter- III. P RELIMINARIES
action Pattern Analysis (UIPA) [29], Reinforcement Learning In this section, we introduce definitions, notations, and
Methods (RLM) [10], and Social Network Analysis (SNA) concepts used in this paper. Key notations are listed in Table
[29], [30]. Considering the limited interpretability of RLM and I.

4 JOURNAL OF IEEE TRANSACTIONS ON ARTIFICIAL INTELLIGENCE, VOL. 00, NO. 0, NOV. 2023

TABLE I
TABLE OF N OTATIONS relationship between each pair of topic and topic aspect, and
x
its weight rik reflects the probability that ui selects an item
whose aspect is cxk ∈ Ĉisub .
Notation Description x
The click probability rik is calculated using Equation 1,
S The recommendation environment. ˆx
where |Ik | denotes the number of ui interacted items whose
U A set of users.
A AI-based algorithm for generating recommendations.
aspects are cxk . In this case, given a belief graph Gi , the sum
x
C A set of predefined topics. of all rik is 1.
ui A user ui in the system.
cx A topic cx
x |Iˆx |
Cxsub A set of aspects associated with a topic cx . rik =P Pk (1)
|Iˆkx′ |
′
cxk An aspect associated with cx . cx′ ∈Ĉi cx
′
∈Ĉxsub
k′ ′
Ikx A set of items whose aspect is cx k. x
im An item. After calculating all click probability rik , the belief degree
Gi Belief network of user ui . bix of user ui towards a topic cx can be formulated as follows:
Vi A set of nodes in Gi .
Ĉi A set of topics that ui interacted with. X
k k
Ĉisub A set of aspects that ui interacted with. bix = − rix log2 (rix ), (2)
Ei A set of edges in Gi . cx sub ∩C sub
k ∈Ĉi x
Eib A set of edges connecting from ui to topics.
Eic A set of edges connecting from topics to aspects. Definition 2. Topic similarity ρ(cx , cy ) represents the similar-
eix An edge from ui to cx . ity between two topics cx and cy , with symmetry ρ(cx , cy ) =
exk An edge from cx to cx k.
bix Belief degree associated on edge eix . ρ(cy , cx ). The similarity measure ρ is defined within the range
x
rik Click probability associated on edge exk . [−1, 1], where a higher value of ρ indicates greater similarity
ρ(cx , cy ) Topic similarity measure between two topics. between the topics.
pi,t Recommendation prompt path for user ui at time step t.
pi,t (k) The kth node in the recommendation prompt path pi,t . As each item is only associated with one topic, we can ob-
f eed A recommendation list comprising f eedoriginal and GI. tain topic embedding by aggregating related item embeddings.
A subset of C that includes only those topics appeared in In this paper, we obtain the topic embedding cx by adopting
Cf eed
the feed.
A list of items suggested by the preference-based recom- the Hadamard product [40] to combine all corresponding
f eedoriginal
mendation system. item embeddings. The similarity between topic cx and cy is
Contextually rich items generated by RGRec, based on an subsequently calculated by the cosine similarity [41]:
GI
explored recommendation prompt path.
A set of items within a f eed accepted by user ui from cx · cy
Mˇi,t
the initial time step to time step t. ρ(cx , cy ) = , (3)
A sequence of recommendation prompt paths shown to ui , ∥cx ∥ ∥cy ∥
P̄i,t where resulting items have been declined over the same
period. where cx and cy denote topic embeddings, cx · cy refers to
the dot product of the topic embeddings, and ∥cx ∥ and ∥cy ∥
denote the corresponding Euclidean norms.
The primary emphasis of this research does not center
A recommendation environment is defined as S = around language embedding. In this paper, our approach
(U, A, C), where U = {u1 , ..., un } represents a set of users, involves the utilization of a pre-trained language model [42]
A refers to an AI-based algorithm for recommending items for the purpose of item embedding.
to users, and C = {c1 , ..., cx } signifies a set of pre-defined Definition 3. A recommendation prompt path pi,t is a
topics. Meanwhile, each cx is associated with a set of aspects sequence of topics explored by a filter bubble-affected user
Cxsub = {cx1 , ..., cxk }, and each cxk is associated with an item ui at a specific time step t, bridging the gap between topics
set Ikx = {i1 , ..., im }. An item can represent a news article or cSP
x
i
(ui strongly preferred) and cLPy
i
(ui less preferred) by
a movie description in real-world applications. To simplify the introducing additional interacted topics. pi,t (k) refers to the
problem, in this paper, we assume each item belongs to only k th topic in a recommendation prompt path pi,t . Furthermore,
one aspect, and each aspect is related to a single topic. the topics within pi,t can be the keywords of prompts for
Definition 1. The belief network of a user ui is repre- the RecomGen module to generate contextually rich items for
sented as a directed graph Gi = (Vi , Ei ). Specifically, recommendation to ui (see Definition 4).
Vi = {ui } ∪ Ĉi ∪ Ĉisub represents a set consisting of three Definition 4. A recommendation list f eed =
distinct types of nodes, where Ĉi ⊂ C denotes a set of topics {f eedoriginal , GI} is a compilation of items recommended
that ui has engaged with, and Ĉisub = {cxk |cxk ∈ Cxsub , cx ∈ to users, where f eedoriginal is comprised of items suggested
Ĉi } represents different aspects related to Ĉi . Meanwhile, by the existing preference-based recommendation system, and
Ei = Eib ∪ Eic represents a composite set of edges, where GI includes contextually rich items generated by RGRec,
Eib = {eix |cx ∈ Ĉi } comprises edges connecting from the user based on an explored recommendation prompt path (refer to
to each topic, and Eic = {exk |cx ∈ Ĉi , cxk ∈ Ĉisub ∩ Cxsub } Definition 3). Given the size of f eed, we introduce a weight
consists of edges connecting from each topic to each of its parameter w to control the proportion of GI within the f eed,
aspects. In Gi , eix denotes ui prefers a topic cx , and the balancing the mix of original and RGRec recommendations.
weight bix associated on eix represents the extent of ui ’s M̌i,t signifies the set of items in a f eed accepted by user ui
belief towards a topic cx . While exk indicates the affiliation from the initial time step to time step t. In contrast, P̄i,t refers

WANG et al.: NUDGING TOWARDS RESPONSIBLE RECOMMENDATIONS 5

provide a recommendation prompt path for the next module,

RecomGen. As depicted in Figure 1, the module takes in two
key inputs from the FBDetect: the user’s preferred “autos”
and the less preferred “news”. The nudging strategy combines
topic similarities to further generate the prompt path based
on these two topics, which is “autos → lifestyle → sports
→ news”. This path acts as an input for the RecomGen.
Finally, the Recommendation Strategy Generation module
(RecomGen) is the “creative” component of RGRec, tasked
with generating an array of items based on the prompt path,
denoted as GI. Drawing on the path charted by the Belief
Nudging module, it constructs recommendations that are rel-
evant and varied, enriching the user’s experience. The GI
blends with the initial recommendation set, ensuring that the
final recommendations delivered to the user are comprehensive
Fig. 1. Overall working process of RGRec for user “U21538”. and engaging. As users engage with these recommendations,
their responses are fed back into the system to refine future
recommendations.
to the sequence of recommendation prompt paths shown to
Through these three core modules, RGRec systematically
ui , where the resulting items have been declined by the user
refines user beliefs and optimal paths in response to user
over the same period.
interactions. Throughout this iterative process, the system con-
Incorporating GI in a f eed bridges the gap between filter
tinuously assesses the likelihood of user acceptance, adapting
bubble-affected users’ preferred and less preferred topics. It
its approach until it aligns with the user’s beliefs, thereby
introduces more interconnected topic-related items, moderates
achieving harmony. Simultaneously, as RGRec expands the
extreme beliefs, and encourages belief harmony.
recommended items using the output of a preference-based
recommendation system, it effectively upholds both responsi-
IV. T HE F RAMEWORK OF R ESPONSIBLE G RAPH - BASED bility and usability.
R ECOMMENDATION
Responsible Graph-based Recommendation (RGRec) is de- A. The Multi-faceted Reasoning-based Filter Bubbles Detec-
signed to guide users gently from a state of information tion module (FBDetect)
imbalance to one of belief harmony. RGRec stands out for
its incorporation of the “nudge” concept and the utilization The Multi-faceted Reasoning-based Filter Bubbles Detec-
of GAI to produce contextually rich items. This approach tion module (FBDetect) is a module designed to identify filter
encourages users to explore interests in topics they may have bubbles and users with extreme beliefs from two perspectives:
originally shown less interest in, providing diverse options. system-level bias and user belief bias. Unlike conventional
Figure 1 illustrates the overview of RGRec’s working process, single-dimensional reasoning models, FBDetect operates in
using a practical example involving the user “U21538”. two complementary modes: Forward Reconnaissance (FR)
As shown in Fig. 1, RGRec operates as a dynamic and and Counter Reconnaissance (CR). These components work
interactive mediator between existing preference-based recom- together to identify and confirm the presence of filter bubbles.
mendation systems and the users, e.g., “U21538” in this ex- The FR component evaluates the recommendation system,
ample. It conforms to the principle of non-coercion, ensuring assessing its potential for causing filter bubbles. It explores
that the user experiences a gradual transition towards belief system-level bias and behaviors to determine if the system
harmony. RGRec comprises three key modules, each integral perpetuates a filter bubble effect. In contrast, the CR com-
to the elimination of filter bubbles. Firstly, the Multi-faceted ponent focuses on users, pinpointing those influenced by
Reasoning-based Filter Bubbles Detection module (FBDe- filter bubbles with extremely imbalanced beliefs. FBDetect
tect) plays a pivotal role in our system by comprehensively offers a comprehensive evaluation of filter bubbles, providing
detecting filter bubbles. Consider the example in Fig. 1, where insights into biased recommendation models and users with
user “U21538” is identified by FBDetect as a filter bubble- extremely imbalanced beliefs. This implementation contributes
affected user with extremely imbalanced beliefs. Specifically, to developing fairer recommendations and interventions to
FBDetect recognizes that this user highly favors “autos” while mitigate the adverse effects of filter bubbles. For a visual
displaying minimal interest in “news”. These precise insights representation of FBDetect’s structure (see Fig. 2).
about users with extremely imbalanced beliefs enable RGRec 1) Forward Reconnaissance (FR) Component: For the FR
to implement targeted interventions to counteract the effects of component, we quantify the potential filter bubble influence
filter bubbles effectively. Subsequently, the FBDetect module of the current recommendation model using mathematical
transfers these findings to the Belief Nudging module, which methods. The aspect coverage score [29] is employed in this
further contributes to achieving user belief harmony. The FR component as a key mathematical validation metric. The
Belief Nudging module forms the core of RGRec. Its main formula for calculating aspect diversity coverage µ is given
task is to use the nudging strategy and topic similarities to below:

6 JOURNAL OF IEEE TRANSACTIONS ON ARTIFICIAL INTELLIGENCE, VOL. 00, NO. 0, NOV. 2023

Forward Reconnaissance Counter Reconnaissance

(FR) (CR)
● Over-specific Detection. User Beliefs Bias autosrearch autossuvs
● Low-diversity Analysis.
User historical cognition
U21538
collection Includes
Includes
6%
Recommendation 6% Includes
autosenthusiasts
belief
logs collection. Includes
6%
belief belief
0.799
16%
0.487 0.445
User belief graph construction shop-holidays
lifestyle 6% Includes
Recommendation autos
logs analysis 6% 6% autosclassics
FBDetect Includes 12%
Includes
User
news 6% 12% Includes
18% Includes
Includes
Includes
Coverage Time-sensitive lifestyledidyouknow autostrucks
Duplicate specific case analysis autossema
autosluxury autossports
Filter bubble- newsus
Filter bubble-
existed
affected user
recommender belief degree less prefer most prefer normal aspect Includes
User detection of topic topic topic preference
YES NO extremely
imbalanced beliefs

Fig. 3. The user belief network for user “U21538”.

Fig. 2. FBDetect: The Multi-faceted Reasoning-based Filter Bubbles Detec-
tion module

|{cx |Ikx ∩f eed̸=∅,∀x,∀k}| |{cx |Ikx ∩f eed̸=∅,∀x,∀k}|

µ= |C| × k
P
|Cxsub |
.
cx ∈C f eed
(4)
The first part measures the diversity of topics by calculating
the proportion of different topics present in the feed relative
to the total number of topics available. While the second part
measures the diversity of the aspects within those topics, where
the proportion of different aspects covered in the feed for the Fig. 4. A specific example of a preference distribution chart within “autos”
topics that are present in the feed, relative to the total number
of aspects in those topics.
2) Counter Reconnaissance (CR) Component: The CR belief networks with the empirical rule (the “68-95-99.7” rule)
component includes two key steps: constructing the user [43] to explore users with extremely imbalanced beliefs.
belief network and detecting users with extremely imbalanced Detecting users with extremely imbalanced beliefs: As
beliefs. previously mentioned, Fig. 3 shows the user belief network
Constructing the user belief network: This step involves of “U21538”. From this figure, we can deduce the user’s
creating a specific belief network for each user, derived from preferences for “autos” and “news”, which are 0.799 and
their historical interaction records. This belief network is 0.445, respectively. To determine whether this user has an
employed to analyze user preferences toward different topics extremely imbalanced belief, we collect statistics on all users’
and identify users with extremely imbalanced beliefs. Fig. 3 preferences for the “autos” and “news”. As illustrated in Fig.
illustrates a representative user belief network for the user 4, these statistics include users’ preferences for “autos”, with
“U21538”, showcasing an example of user belief network the x-axis representing the preference value and the y-axis
construction within the CR component. representing the number of users. We identify “U21538” as
In Fig. 3, the red, blue, and yellow nodes represent the topics a user with extremely imbalanced beliefs by following these
“news”, “autos”, and “lifestyle”, with which user “U21538” steps:
has historically interacted. The pink nodes represent the as- • Verify Normal Distribution: The initial step involves ver-
pects related to each topic that the user has interacted with. The ifying whether the preference distribution chart conforms
weights over edges labeled “Includes”, linking a topic to its to a normal distribution. To achieve this, we utilize the
aspects, are calculated based on the user’s click probabilities Kolmogorov-Smirnov (K-S) test [44]. In our scenario, the
towards these aspects. Additionally, the connection between K-S test is a statistical method employed to compare the
the user and a topic, “belief”, indicates the user’s preference preference distribution of a specific topic with the normal
towards the topic and is calculated using the entropy metric distribution. The results of the K-S test typically include
in Equation 2. the value of the K-S statistic and its associated p-value.
The CR component constructs a specific user belief network The p-value denotes the probability of observing the K-
for each user in our dataset, enabling clear identification of S statistic under the assumption that the distributions of
users’ preferences. Its primary aim is to identify users with the two datasets are identical. A small p-value (in this
extremely imbalanced beliefs. We combine these constructed study, we set the threshold at 0.05) allows us to reject

WANG et al.: NUDGING TOWARDS RESPONSIBLE RECOMMENDATIONS 7

the null hypothesis that the two datasets share the same user ui at time step t. This path incorporates contextually rich
distribution, indicating statistically significant differences. items into the recommendations, thus providing more diverse
Therefore, in Fig. 4, the p-value exceeds 0.05, suggesting content and promoting belief harmony.
that the preference distribution within “autos” conforms In the recommendation prompt path pi,t for the user ui ,
to a normal distribution. cSP
x
i
and cLP
y
i
represent the user’s most and less preferred
• Distribution Segment: Once it is confirmed that the pref- topics, respectively, serving as the path’s start and end points.
erence distribution chart adheres to a normal distribution, The initial node of the path is denoted as pi,t (k) = cSP
x
i
with
th
the next step involves segmenting the distribution using k=1. The selection of the subsequent node, the (k + 1) topic
the “68-95-99.7” rule [43], as depicted in Fig. 4. In the cx from the set C, is determined by maximizing the following
figure, it can be seen that approximately 68% of the expression:
data falls within one standard deviation of the mean,
about 95% within two standard deviations, and roughly
pi,t (k + 1) = arg max ρ(cSP
x , cx′ ) + bix′ ∗ rejw.t ,
i
(5)
99.7% within three standard deviations. In our study, we cx′
categorize users outside the 68% area (outside the yellow
where ρ(cSP SPi
x , cx′ ) calculates the topic similarity between cx
i
lines) as having extreme beliefs in the topic. Therefore,
and cx′ . The term bix′ is the belief degree of user ui towards
when user U21538 (see Fig. 3) prefers “autos” with 0.799
topic cx′ at the current time step, rejw.t is a rejection weight,
degrees, falling outside the 68% area in Fig. 4, and shows
typically set to 1.
less interest in “news” with 0.445 degrees, also falling
We incorporate a tolerance threshold, denoted as θ, to
outsides the 68% area in the “news” preference distri-
modulate the parameter rejw.t , emphasizing the significance
bution chart. This user is identified as having extremely
of user feedback. Users impacted by filter bubbles tend to
imbalanced beliefs and is further forwarded to the Belief
have their decisions heavily influenced by these bubbles. An
Nudging module to mitigate the extremely imbalanced
item congruent with a user’s beliefs is more likely to be
beliefs, i.e., filter bubbles.
accepted, whereas items not aligned with preferred topics often
face acceptance challenges. To evaluate the effectiveness of
B. Belief Nudging Module a topic within a recommendation prompt path, we monitor
RGRec combines users’ belief graphs and nudging tech- the frequency of rejections for topic-related items. If a user
niques to gently stimulate users’ interests in topics that were consistently rejects items from GI that are generated based
initially less preferred. The primary goal of the Belief Nudging on the recommendation prompt path, these items are deemed
module within the RGRec framework is to identify the most ineffective, leading to an assignment of a rejw.t value of -1.
effective recommendation prompt path for the subsequent This approach adapts effectively to evolving user preferences.
RecomGen. This is achieved by bridging the gap between the The exploration process concludes once the path pi,t includes
user’s most favored and less preferred topics. the user’s less preferred topic cLP
y .
i

1) Adaptive Path Exploration Algorithms: The recommen- Utilizing the start and end nodes identified as “autos” and
dation prompt path begins with topics that users highly favor “news” in Fig. 5, and employing the adaptive path exploration
(e.g., “autos” as shown in Fig. 3) and ends with those topics algorithm, the recommendation prompt path at time t is
they are less inclined towards (such as “news” in Fig. 3). To formulated as pi,t = “autos → lifestyle → sports → news”.
connect these points, an adaptive path exploration algorithm 2) Nudging Strategy: Our nudging process incorporates in-
discovers additional topics, forming a comprehensive recom- cremental computing techniques [46] to enhance the efficiency
mendation prompt path and laying the groundwork for future of recommendation calculations. By breaking down the recom-
nudge-based recommendations. mendation path into smaller segments, or sub-paths, the system
The nudge recommendation strategy in RGRec is based on recalibrates only the affected sub-path in response to a user’s
this recommendation prompt path, where the adaptive path specific preference, instead of recalculating the entire path.
exploration algorithm automatically constructs the path based This segmented approach is particularly effective for managing
on user feedback. This forms a closed-loop feedback system lengthy paths and surpasses the capabilities of traditional
that interlinks user feedback and system recommendations, dy- sequential recommendation systems. It not only reduces the
namically adapting to deliver contextually appropriate prompts number of recommendations needed but also increases overall
tailored to the user’s current preferences. efficiency. Once the path “autos → lifestyle → sports →
The adaptive path exploration algorithm, inspired by the news” is established, the nudging strategy is employed to
shortest path exploration algorithm known as Dijkstra’s algo- refine and finalize the recommendation prompt path. For a
rithm [45], starts from a central point, traverses neighboring clear illustration, both Figure 5 and Algorithm 1 demonstrate
points, and identifies the point with the highest weight as the the recommendation process within the RKGRec nudging
starting point for the next step. We have proposed an enhanced framework. This process includes generating recommendation
version of this algorithm to accommodate the dynamic nature paths through nudging, creating GI by the RecomGen, and
of user and topic relationships. This algorithm integrates the subsequently reconstructing the user belief graph.
evolving user perceptions and topic relationships into the path In Fig. 5, the identified optimal path is “autos → lifestyle
discovery process. The objective is to discover a recommen- → sports → news”. This sequence forms the basis for the
dation prompt path pi,t for an identified filter bubble-affected nudging strategy, which is fine-tuned based on user feedback.

8 JOURNAL OF IEEE TRANSACTIONS ON ARTIFICIAL INTELLIGENCE, VOL. 00, NO. 0, NOV. 2023

optimal path start: autos lifestyle sports news :end Algorithm 1 Nudge Recommendation Process of RGRec
prompt path: autos lifestyle Input: the current recommendation prompt path
round 1
RecomGen pi,t = {cSP
x
i
, ..., cLP
y
i
} as the current prompt p
restart interacted Title
Title
non-interacted restart
GI
Title

user beliefs 1: Initialise an empty heap Q = []

nodes changed nodes unchanged restart 2: Q = Binary Split Function(p) ▷ Binary Split Function adopts
restart
new path sports news autos incremental computing techniques to break down path p
round 2
RecomGen RecomGen 3: while length(Q) > 0 do
interacted
interacted non-interacted
non-interacted
Title
Title

4: Set current prompt p = Q[0]

GI
Title Title
Title

GI
Title

restart 5: Generate contextually rich items GI with prompt p = Q[0]

user beliefs new path
user beliefs new path
6: Recommend GI to user ui
nodes unchanged nodes changed unchanged nodes changed
7: if ui accepts GI then
round 3 news new path
autos lifestyle new path
8: M̌i,t = M̌i,t ∪ GI ▷ M̌i,t represents user ui ’s collection
RecomGen
Title
Title
RecomGen of accepted items until time step t
interacted GI non-interacted Update user belief graph Gui
Title

interacted GI non-interacted
Title
Title

Title 9:
user beliefs new path
user beliefs new path
10: Q.pop(0)
11: else
12: P̄i,t .append(p) ▷ P̄i,t records user ui ’s declined
recommendation prompt paths from time 0 to t
Fig. 5. An actual example of nudging recommendations process 13: if count(P̄i,t , p) > θ then
14: Update rejection weight rejw.t
15: end if
16: p1t , p2t = Binary Split Function (p)
Conforming to the principles of incremental computing, the 17: Q.pop(0)
primary focus is initially on the “autos → lifestyle” segment. 18: Push p1t , p2t to Q
RGRec dynamically updates the user’s belief network as the 19: end if
user positively engages with the “GI” content generated from 20: if Binary Split Function (p) is None then
RecomGen based on this segment. If the user’s most and 21: Reschedule path
22: end if
least preferred topics remain constant, the model progresses to 23: end while
recommend the subsequent segment, “sports → news”, instead 24: function B INARY S PLIT F UNCTION(pt )
of restarting the entire path exploration process. However, 25: if length(pt ) == 2 then
if the user’s topic preferences shift, these altered topics are 26: return
reintroduced to the adaptive path exploration algorithm to 27: else
28: if mod(length(pt ), 2) == 1 then
create a new recommendation path. 29: p1t = pt [0 : length(p t )−1
]
2
2 length(pt )+1
30: pt = pt [ 2
:]
C. The Generative Artificial Intelligence-based Recommenda- 31: else
tion Strategy Generation module (RecomGen) 32: p1t = pt [0 : length(p
2
t)
]
2 length(pt )
33: pt = pt [ 2
:]
When combined with nudge recommendation techniques, 34: end if
the RecomGen efficiently exploits the interconnectedness of 35: end if
information. This combination presents a solution to address 36: return p1t , p2t
information gaps that may arise during end-to-end recom- 37: end function
mendation processes. By employing the capabilities of the
Large Language Models (LLM), e.g., GPT-3.5 Turbo, this
approach offers rich semantic information at each step in the semantic information provided by RecomGen, and employs a
recommendation path, thereby strengthening the relationships nudge strategy to establish strong connections between data
between individual points. This strategy effectively engages points. This strategy effectively stimulates user interest in
users’ interest in specific topics, fostering intrigue in areas topics that initially receive less attention and promotes user
they might find less attractive. belief harmony.
As previously mentioned, a nudging prompt pi,t = Acceptance Probability Equation
{cxSPi , ..., cLP
y } is generated for each time step. This prompt
i The RecomGen incorporates a list of contextual rich items
represents an optimal path between the starting node cSP x ,
i GI into the original recommendation list f eedoriginal and
and the end node cLP y
i
. To leverage rich contextual information generates the final recommendation list f eed = {i1 , i2 , .., ij }.
i
from point-to-point paths within the vast landscape of big data, Once user ui receives this f eed, the probability APuji that
these paths are inputted into the RecomGen as keywords or whether to accept an item ij in the recommendation list can
prompts. This generates contextually rich items GI based on be calculated using Equation 6:
the prompt path.
bix
For example, in Fig. 5, the responsibility of the RecomGen APuiji = P , ij ∈ Ikx (6)
is to obtain the path from each f eed and generate GI based cx b ′
′ ∈Ĉi ix

on the path. In conclusion, integrating the RecomGen with the

Pbix represents the belief degree of ui towards topic cx ,
where
nudge strategy in RGRec effectively utilizes interconnections. and c ′ ∈Ĉi bix′ calculates the total belief degrees of ui .
x
This approach bridges information gaps, leverages the rich

WANG et al.: NUDGING TOWARDS RESPONSIBLE RECOMMENDATIONS 9

V. E XPERIMENTS •Neural Graph Collaborative Filtering (NGCF) [51]: En-

We conduct experiments from two general directions: sys- hances recommendation by using user-item graphs to
tem and user. The system-centered experiment primarily aims model collaborative signals.
to demonstrate the effectiveness of RGRec as an intermediate • LightGCN (LGCN) [52]: A simplified, yet effective,

agency in alleviating the system filter bubble. From the user’s model focusing on neighborhood aggregation; outper-
perspective, four user-centered experiments are conducted. forms NGCF while being easier to train.
These include detecting RGRec’s positive effect on increasing • Disentangled Graph Collaborative Filtering (DGCF) [53]:

user belief diversity, examining its effectiveness in motivat- Captures user intent diversity by analyzing user-item
ing filter bubble-impacted users’ interests in topics they are relationships.
initially less interested in, and analyzing RGRec’s ability to • RGRec: This method suggests a set of items (designated

reduce the number of filter bubble-impacted users. Finally, as GI) as recommendation feeds, derived using the
we perform two parametric analyses to assess the effects of RGRec approach.
different RGRec recommendation weights and explore the Building upon the baselines described above, in our experi-
impact of the threshold in RGRec on user belief diversity. ments, we combined RGRec with each of the baselines (except
the standalone RGRec), producing six additional experimental
baselines. These combinations are all named with the super-
A. Experiment Setup
script “∗”, such as CB∗ and LGCN∗ . In total, we used 11
1) Dataset: In the experiments, we utilize two real-world baselines for comparative experiments, aiming to demonstrate
datasets: the Microsoft News Dataset (MIND) 1 and IMDB the superior performance of the models when combined with
2
Dataset. MIND is a public news recommendation dataset RGRec in breaking the filter bubble and moderating user
encompassing user interaction data gathered from Microsoft extreme beliefs.
News. It comprises data from 5,000 users, encompassing Evaluation Metrics:
230,117 user reading records and 51,287 news with 17 topics. We assess RGRec from both system and user perspectives.
IMDB is a movie recommendation dataset consisting of 25,000 From the system viewpoint, Experiment V-C1 measures rec-
movie rating records from 333 users and a selection of 2,586 ommendation diversity using aspect coverage µ as defined in
movies in 16 movie topics. We adopt a pre-trained language Equation 4. To confirm that observed differences in experimen-
model, BERT [42], to represent textual features in a vector tal outcomes are not due to chance, we employ the Two-Tailed
space, capturing the semantic essence of the items. Such a Test t and corresponding p-value [54], attributing significant
method has already learned much about language structures differences to model factors. The value of t is shown below:
and patterns [47] and has been widely adopted in recommen-
dation studies [48], [49], [50]. X̄1 − X̄2
t= q , (7)
The FBDetect module in RGRec, identified 180 and 20 filter Sp n11 + n12
bubble-affected users from the MIND and IMDB datasets,
respectively. These identified users, denoted as uF B in our where X̄1 and X̄2 are the means of the two sample groups,
paper, are the focus of subsequent filter bubble mitigation and Sp is the pooled standard deviation. n1 and n2 represent
experiments. the sample sizes of these groups. The p value represents the
2) Simulation of User Behaviors.: Given the impracticality probability of observing the t statistic observed, assuming no
and high cost associated with online testing for researchers, we significant difference exists between the two sample groups.
have designed an offline evaluation approach: (1) Implement From the user perspective, our experiments focus on three
an “Acceptance Probability Equation” 6 to simulate user feed- areas. In Experiment V-C1, we evaluate the diversity of users’
back, (2) generate recommendations using a “Nudge Strategy” belief networks using the same µ metric in Equation 4.
based on the simulated user feedback, and (3) evaluate the Experiment V-C2 analyzes the evolution of users’ interests in
recommendations by considering diversity and efficacy. highly favored and less favored topics over time, employing
Equation 2. Lastly, Experiment V-C3 involves counting the
B. Parameter settings and baselines number of filter bubble-affected users, using the CR of the
FBDetect module, as detailed in Section IV-A2.
Baselines: We assess the performance of RGRec in com- Parameters: The proportion w of RGRec-generated items
parison with several established baseline methods: GI is considered as a parameter in our experiments. We
• Content-Based Filtering (CB) [18]: This strategy rec- analyze the impact of varying proportions w within a rec-
ommends existing items based solely on content-based ommendation f eed on user belief and the diversity of the
filtering. recommendation system. Additionally, we conduct a param-
• Collaborative Filtering (CF) [17]: This approach rec- eter analysis experiment of the tolerance threshold θ, which
ommends existing items using only user collaborative involves tracking user feedback regarding the generated GI,
filtering. described in Section IV-B.
1 https://msnews.github.io/
2 https://www.kaggle.com/datasets/meastanmay/imdb- C. Experimental Results
dataset?select=tmdb 5000 movies.csv/ 1) Experiment 1: Coverage Analysis: The primary aim
of this experiment is to assess the impact of RGRec on

10 JOURNAL OF IEEE TRANSACTIONS ON ARTIFICIAL INTELLIGENCE, VOL. 00, NO. 0, NOV. 2023

TABLE II
C OVERAGE A NALYSIS OF R ECOMMENDATION M ODELS BASED ON MIND AND IMDB DATASETS . B OLDFACE DENOTES THE HIGHEST SCORE . M ARKING
WITH UNDERLINE DENOTES THE SIGNIFICANCE P - VALUE<0.05 COMPARED WITH THE BASE MODEL .

MIND IMDB
Times
CB CB∗ CF CF∗ DGCF DGCF∗ NGCF NGCF∗ LGCN LGCN∗ RGRec CB CB∗ CF CF∗ DGCF DGCF∗ NGCF NGCF∗ LGCN LGCN∗ RGRec
f eed1 0.058 0.294 0.184 0.335 0.176 0.294 0.294 0.470 0.294 0.412 0.294 0.062 0.500 0.212 0.526 0.188 0.438 0.250 0.375 0.125 0.479 0.500
f eed2 0.059 0.205 0.186 0.202 0.176 0.353 0.235 0.470 0.235 0.412 0.195 0.062 0.345 0.207 0.386 0.125 0.500 0.250 0.563 0.250 0.354 0.344
f eed3 0.058 0.154 0.181 0.207 0.176 0.176 0.294 0.294 0.294 0.294 0.132 0.062 0.238 0.209 0.328 0.125 0.469 0.250 0.375 0.125 0.250 0.216
f eed4 0.059 0.145 0.179 0.241 0.117 0.235 0.294 0.294 0.294 0.264 0.123 0.062 0.180 0.201 0.274 0.125 0.313 0.219 0.281 0.125 0.188 0.147
f eed5 0.059 0.151 0.182 0.215 0.117 0.117 0.294 0.294 0.294 0.294 0.119 0.062 0.163 0.207 0.259 0.281 0.344 0.219 0.281 0.188 0.250 0.127
f eed6 0.059 0.199 0.184 0.242 0.059 0.117 0.294 0.294 0.235 0.265 0.171 0.062 0.163 0.216 0.254 0.250 0.281 0.188 0.313 0.125 0.188 0.127
f eed7 0.059 0.205 0.176 0.212 0.117 0.117 0.235 0.412 0.294 0.294 0.174 0.062 0.190 0.209 0.268 0.188 0.250 0.219 0.344 0.125 0.250 0.149
f eed8 0.059 0.178 0.186 0.215 0.176 0.117 0.235 0.353 0.235 0.324 0.154 0.062 0.184 0.197 0.269 0.125 0.156 0.219 0.219 0.250 0.250 0.158
f eed9 0.059 0.164 0.185 0.224 0.117 0.117 0.294 0.471 0.294 0.382 0.139 0.062 0.170 0.214 0.247 0.062 0.188 0.188 0.250 0.125 0.188 0.139
f eed10 0.059 0.156 0.191 0.221 0.117 0.117 0.294 0.412 0.235 0.352 0.136 0.062 0.164 0.218 0.243 0.125 0.125 0.219 0.219 0.125 0.1875 0.135
sum. 0.588 1.851 1.832 2.315 1.352 1.941 2.765 3.765 2.706 3.294 1.501 0.625 2.297 2.089 3.053 1.594 3.063 2.219 3.219 1.563 2.583 2.041
Improv. - 214.79% - 26.31% - 43.48% - 36.17% - 21.74% - - 267.6% - 46.15% - 92.16% - 45.07% - 65.33% -

content diversification compared to the baseline model, and into recommendation models significantly enhances the diver-
its effect on user belief networks. To accomplish this, we sity of content recommended to users, which is crucial for
identified users affected by filter bubbles using specific criteria, mitigating filter bubble effects.
including users such as “U25354” from the MIND dataset Coverage Analysis for User Belief Networks. Table III
and “U128” from the IMDB dataset. For each of these users, offers a detailed examination of the impact of different rec-
we conducted a series of experimental rounds where they ommendation models on the diversity of user beliefs across
received 10 recommendations in each round, corresponding the MIND and IMDB datasets.
to a unique “feed”. This process was repeated for 100 rounds. In both datasets, models integrated with RGRec demonstrate
The repetition of these rounds aimed to ensure the reliability of a significant improvement in total diversity coverage (sum.)
our results by reducing the influence of randomness. Finally, compared to their respective baseline models. This indicates
the data from these 100 rounds were averaged to obtain the that integrating RGRec effectively broadens the range of user
final experimental results, offering a comprehensive evaluation beliefs. The most notable diversity gains are seen in models
of RGRec’s effectiveness in enhancing diversity in recommen- like DGCF∗ , NGCF∗ , and LGCN∗ , especially in the MIND
dations. dataset. This highlights the effectiveness of these RGRec-
Coverage Analysis for Systems: Initially, we analyzed the enhanced models in diversifying user recommendations.
evolution of content diversity across seven different models, The Improv. metric shows a marked percentage increase
focusing on selected users (user “U25354” from MIND and in user belief diversity for the RGRec integrated models,
user “U128” from IMDB). The detailed results are presented particularly in the MIND dataset. For instance, the DGCF∗
in Table II, showcasing each model’s diversity coverage. Note model shows improvements exceeding 600%. The consistent
that “sum.” represents each model’s total diversity coverage enhancements across various models and both datasets rein-
degree throughout the recommendation process, and “Improv.” force RGRec’s effectiveness in enhancing user belief diversity.
indicates the growth rate in diversity. The statistical significance values for all models integrated
In the MIND dataset, the models with wR suffix, indicating with RGRec on both datasets are all below the 0.05 threshold
RGRec integration, generally exhibit significant improvements compared with the base model, confirming that the improve-
in diversity coverage compared to their standard counterparts. ments in user belief diversity are statistically significant and
Notably, models such as DGCF∗ , NGCF∗ , and LGCN∗ show attributable to the model variations, particularly due to RGRec
substantial enhancements, as reflected by their high sum. integration. This statistical robustness emphasizes the reliabil-
values and growth rates. This indicates that incorporating ity of the trends observed.
RGRec markedly enhances recommendation diversity. Overall, the analysis confirms that the inclusion of RGRec
Similarly, in the IMDB dataset, models integrated with markedly enhances the diversity of user beliefs, which is
RGRec demonstrate considerable improvements in diversity. crucial in mitigating filter bubble effects and enriching user
The impact of RGRec is particularly pronounced in models belief diversity.
like NGCF∗ and LGCN∗ , which display the highest growth 2) Experiment 2: User Beliefs Analysis: RGRec is designed
rates in diversity. to present diverse information to users based on existing
The statistical significance values for all models integrated preference-based recommendation systems. Our goal is to
with RGRec on both datasets are all below the 0.05 threshold demonstrate that users may come to accept information they
compared with the base model, suggesting that the improve- initially preferred less through RGRec recommendations over
ments in diversity coverage by RGRec-integrated models are a relatively short timeframe, thereby broadening their per-
statistically significant compared to their counterparts. This spectives. To this end, we selected user “U276” from the
underscores the vital role of RGRec in augmenting recom- IMDB dataset, who presents a shorter recommendation path
mendation diversity. pshorter.u , indicating closer feature distances between most and
Overall, the analysis demonstrates that integrating RGRec less interested topics. This approach allows us to observe
changes in user beliefs within a short period. Additionally,

WANG et al.: NUDGING TOWARDS RESPONSIBLE RECOMMENDATIONS 11

TABLE III
C OVERAGE ANALYSIS OF USER BELIEFS ON THE MIND AND IMDB DATASETS . B OLDFACE DENOTES THE HIGHEST SCORE . M ARKING WITH UNDERLINE
DENOTES THE SIGNIFICANCE P - VALUE<0.05 COMPARED WITH THE BASE MODEL .

MIND IMDB
Times
CB CB∗ CF CF∗ DGCF DGCF∗ NGCF NGCF∗ LGCN LGCN∗ RGRec CB CB∗ CF CF∗ DGCF DGCF∗ NGCF NGCF∗ LGCN LGCN∗ RGRec
f eed1 0.056 0.090 0.056 0.104 0.059 0.294 0.059 0.294 0.059 0.294 0.047 0.038 0.097 0.049 0.090 0.062 0.375 0.031 0.375 0.062 0.062 0.085
f eed2 0.059 0.116 0.075 0.133 0.059 0.471 0.059 0.294 0.059 0.353 0.074 0.054 0.147 0.082 0.159 0.062 0.375 0.094 0.375 0.062 0.062 0.104
f eed3 0.059 0.126 0.089 0.161 0.059 0.471 0.059 0.294 0.059 0.353 0.091 0.059 0.172 0.111 0.198 0.062 0.375 0.094 0.375 0.062 0.062 0.126
f eed4 0.059 0.135 0.099 0.186 0.059 0.471 0.059 0.294 0.059 0.353 0.106 0.062 0.183 0.133 0.223 0.125 0.375 0.094 0.375 0.062 0.125 0.145
f eed5 0.059 0.144 0.110 0.204 0.059 0.471 0.059 0.294 0.059 0.353 0.116 0.062 0.191 0.151 0.243 0.125 0.375 0.125 0.375 0.062 0.125 0.159
f eed6 0.059 0.158 0.122 0.221 0.059 0.471 0.059 0.294 0.059 0.353 0.132 0.062 0.197 0.172 0.257 0.125 0.375 0.125 0.375 0.062 0.125 0.164
f eed7 0.059 0.165 0.131 0.228 0.059 0.471 0.059 0.294 0.059 0.353 0.142 0.062 0.208 0.189 0.279 0.125 0.375 0.125 0.375 0.062 0.125 0.174
f eed8 0.059 0.171 0.138 0.238 0.059 0.471 0.059 0.353 0.059 0.393 0.148 0.062 0.219 0.203 0.286 0.125 0.375 0.125 0.375 0.062 0.125 0.181
f eed9 0.059 0.178 0.149 0.245 0.059 0.471 0.059 0.353 0.059 0.393 0.155 0.062 0.228 0.213 0.300 0.125 0.375 0.125 0.375 0.062 0.125 0.188
f eed10 0.059 0.184 0.160 0.252 0.059 0.471 0.059 0.353 0.059 0.393 0.167 0.062 0.233 0.220 0.310 0.125 0.375 0.125 0.375 0.062 0.125 0.192
sum. 0.585 1.467 1.128 1.970 0.588 4.529 0.588 3.112 0.590 3.588 1.178 0.587 1.873 1.522 2.344 1.063 3.750 1.063 3.750 0.625 1.062 1.517
Improv. - 150.75% - 74.56% - 669.99% - 430.00% - 510.00% - - 218.93% - 53.98% - 252.94% - 252.94% - 70.00% -

3) Experiment 3: Filter Bubble Users Detection: Table IV

provides a comprehensive overview of the filter bubble effect
across different recommendation models in both the MIND
and IMDB datasets. This experiment focuses on tracking the
changes over time in the number of users influenced by the
filter bubble effect.
In the MIND dataset, a consistent pattern is observed where
(a) Temporal Variation in User Beliefs (b) Temporal Variation in User Be-
about Topics of Most Interest and Less liefs about Topics of Most Interest and models integrated with RGRec consistently affect fewer users
Interest on MIND Less Interest on IMDB with the filter bubble than their original counterparts. This
trend is evident across all recommendation feeds, showcasing
Fig. 6. Temporal Variation in User Beliefs about Topics of Most Interest RGRec’s effectiveness in reducing the filter bubble impact.
and Less Interest
Particularly, the original DGCF and NGCF models, which
initially show a high number of users affected, see a significant
user “U18469”, with a longer recommendation path plonger.u , decrease following RGRec integration. This highlights the
is included in the experiment to represent cases with longer model’s capability to diversify user recommendations.
recommendation paths. Similarly, in the IMDB dataset, RGRec-enhanced models
This experiment employs the CF∗ model to investigate also demonstrate a reduction in the number of users affected by
whether users, in the context of RGRec, will shift their the filter bubble. For example, the CF∗ model shows a decrease
focus from their highly preferred topic to a topic they are from 6 to 4 users affected and then exhibits fluctuation,
less interested in. Fig. 6 illustrates the temporal evolution of reflecting the dynamic nature of user preferences and the
interest levels in topics initially less interesting and highly system’s adaptability. RGRec’s ability to adjust the recommen-
interesting to the users, as indicated by their belief networks. dation path based on user feedback likely contributes to these
For the user with longer paths, “plonger ”, we have adjusted changes, allowing for a more responsive recommendation
the timescale in Fig. 6 so that a single time interval now rep- system that aligns with evolving user interests.
resents 10 recommendation feeds. This allows us to measure Overall, the analysis from the table confirms the effective-
the user’s interest level in the “autos” and “video” every 10 ness of the RGRec strategy in mitigating the filter bubble effect
recommendations. In contrast, for the user with shorter paths, across different recommendation models and datasets. The
“pshorter ”, we track belief changes after each recommendation. observed temporal fluctuations further highlight the dynamic
In these figures, “autos” and “Action” refer to the topics of interaction between users and the recommendation system.
high interest, while “travel” and “Drama” represent the less 4) Experiment 4: Parameter Analysis-1: As mentioned
favored topics for each user. in Subsection V-B, RGRec utilizes two key parameters: the
From both figures, it is observed that the users’ preferences proportion w of RGRec-generated contextually rich items GI,
for “autos” and “Action” remain relatively stable or exhibit and a tolerance threshold θ that tracks user feedback on these
a slight decline over time. On the other hand, continuous items. This experiment examines the impact of varying the
recommendations progressively increase their interest in the weights w of RGRec-generated items within a recommenda-
initially less favored topics, such as “video” and “Drama”. tion feed on user belief diversity. We focus on how changing
This trend is particularly pronounced for the IMDB user, who the weight of RGRec-generated items influences the diversity
shows a significant increase in interest in “Drama”. These of user perspectives. By adjusting the w parameter, we as-
observations robustly validate the effectiveness of the RGRec- sess the extent of RGRec’s influence on the recommendation
based model in promoting belief harmony among users by outcomes. The experiment employs user “U1629” from the
diversifying their interests. MIND dataset as a case study and uses the CF∗ model.
As shown in Fig. 7, the weight assigned to RGRec recom-
mendations has a significant effect on user belief diversity.

12 JOURNAL OF IEEE TRANSACTIONS ON ARTIFICIAL INTELLIGENCE, VOL. 00, NO. 0, NOV. 2023

TABLE IV
F ILTER B UBBLE U SERS D ETECTION ON MIND AND IMDB DATASETS

MIND IMDB
Times
CB CB∗ CF CF∗ DGCF DGCF∗ NGCF NGCF∗ LGCN LGCN∗ RGRec CB CB∗ CF CF∗ DGCF DGCF∗ NGCF NGCF∗ LGCN LGCN∗ RGRec
f eed1 2 2 28 24 ↓ 26 18 ↓ 26 16↓ 20 16 ↓ 26 0 0 6 4↓ 6 5 ↓ 6 5 ↓ 6 5 ↓ 4
f eed2 2 2 28 24 ↓ 26 16 ↓ 26 16↓ 20 16 ↓ 26 0 0 6 4↓ 6 5 ↓ 6 5 ↓ 6 5 ↓ 4
f eed3 2 2 28 24 ↓ 26 14 ↓ 26 18↓ 20 18 ↓ 26 1 0↓ 6 4↓ 6 5 ↓ 8 5 ↓ 7 5 ↓ 4
f eed4 4 4 28 24 ↓ 26 14 ↓ 28 18↓ 22 18 ↓ 26 2 0↓ 6 6 8 5 ↓ 8 5 ↓ 7 5 ↓ 4
f eed5 4 4 28 24 ↓ 26 10 ↓ 28 18↓ 24 20 ↓ 26 3 1↓ 6 6 8 5 ↓ 8 5 ↓ 7 5 ↓ 4
f eed6 4 4 28 24 ↓ 26 10 ↓ 28 18↓ 26 18 ↓ 26 3 2↓ 6 6 8 5 ↓ 8 5 ↓ 7 5 ↓ 4
f eed7 4 4 28 24 ↓ 28 10 ↓ 28 18↓ 26 18 ↓ 26 3 3 6 6 8 5 ↓ 8 5 ↓ 8 5 ↓ 4
f eed8 6 4 ↓ 28 24 ↓ 28 10 ↓ 28 16↓ 28 18 ↓ 24 3 2↓ 6 6 8 5 ↓ 8 5 ↓ 8 5 ↓ 4
f eed9 6 4 ↓ 28 24 ↓ 28 10 ↓ 28 16↓ 28 18 ↓ 24 3 2↓ 6 6 8 5 ↓ 8 5 ↓ 8 5 ↓ 4
f eed10 10 4 ↓ 28 24 ↓ 28 10 ↓ 28 16↓ 28 16 ↓ 24 4 2↓ 6 5↓ 8 5 ↓ 8 5 ↓ 8 5 ↓ 4

User Beliefs Diversity Change on Different Nudge Weights User Beliefs Diversity Change on Different Tolerance Threshold
feed10 feed10
feed9 feed9
feed8 feed8
feed7 feed7
feed6 feed6

Feeds
Feeds

feed5 feed5
feed4 feed4
w=0
feed3 w=0.2 feed3
feed2 w=0.4
w=0.6 feed2 =1
feed1 w=0.8 =2
w=1 feed1 =3
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.00 0.05 0.10 0.15 0.20 0.25
Coverage Coverage

Fig. 7. User Beliefs Diversity Change on Different Nudge Weights Fig. 8. User Beliefs Diversity Change on Different Tolerance Threshold

It becomes clear that as the number of recommendations thresholds result in greater user belief diversity, showcasing
increases over time, the impact of the recommendation weight the system’s enhanced adaptability to user feedback. Notably,
grows more pronounced. Notably, when the weight is set at when the threshold value is 3, the user belief diversity achieves
0.4, users are more likely to accept the recommendations, and the largest value, surpassing 0.25. This implies that a moderate
this influence remains steady. However, after a certain num- threshold encourages the system to adjust its recommendations
ber of recommendations, the impact of RGRec-recommended based on user feedback while still maintaining substantial di-
information on users reaches its peak. The more significant versity. The experiment intentionally avoids setting thresholds
the fraction of items recommended by RGRec in the total higher than 3 to prevent persistently recommending items
recommendation list, the more significant the impact on users that users dislike, which could lead to disengagement from
in the later stages of recommendations. the recommendation system. Across all thresholds, initial
In conclusion, the analysis shows that the weight of RGRec acceptance of recommendations is typically low. However,
recommendations considerably affects user belief diversity. user acceptance of RGRec gradually increases as it adapts to
The influence of this weight amplifies as the number of rec- feedback from multiple rejections, refining its nudge strategy
ommendations increases. However, there is a saturation point and enhancing the overall effectiveness of the approach.
at which the impact of RGRec-recommended information on In conclusion, this experiment highlights the critical role of
users reaches its maximum. In the experiment, we select threshold settings in influencing user belief diversity. The find-
w=0.6 as the RGRec recommendation weight owing to its ings suggest that a threshold value of 3 maximizes user belief
consistently improving performance. diversity while appropriately respecting user preferences.
5) Experiment 4: Parameter Analysis-2: In this experiment,
we investigate the impact of the tolerance threshold in RGRec VI. D ISCUSSION
on user belief diversity. The threshold setting reflects the The experiments illustrate that models integrating RGRec
model’s consideration of user feedback, determining how as an intermediary significantly outperform those without
many times a user must reject recommendations before RGRec RGRec. RGRec is particularly potent in reducing filter bubbles
alters its nudge recommendation strategy. We simulate various and achieving a more balanced user belief network. It proves
scenarios where users reject recommendations to mimic dif- to be more effective and efficient in diversifying recommen-
ferent feedback situations. dations and fostering increased user belief diversity as the
Analyzing Fig. 8, it is clear that user belief diversity recommendation count grows.
gradually increases as the number of recommendations grows, Furthermore, the experiments assessed the impact of RGRec
with different threshold values playing a pivotal role. Higher on users with varying weights of RGRec recommendations.

WANG et al.: NUDGING TOWARDS RESPONSIBLE RECOMMENDATIONS 13

The findings reveal its effectiveness in increasing users’ ac- [3] A. G. Ekström, D. C. Niehorster, and E. J. Olsson, “Self-imposed filter
ceptance of initially less preferred information, particularly bubbles: Selective attention and exposure in online search,” Computers
in Human Behavior Reports, vol. 7, p. 100226, 2022.
for those with shorter recommendation paths. For users with [4] M. Elahi, D. Jannach, L. Skjærven, E. Knudsen, H. Sjøvaag, K. Tolonen,
longer paths, the model aids in gradually adapting to less Ø. Holmstad, I. Pipkin, E. Throndsen, A. Stenbom et al., “Towards
favored information, underscoring its capacity to improve responsible media recommendation,” AI and Ethics, pp. 1–12, 2022.
[5] F. Alatawi, L. Cheng, A. Tahir, M. Karami, B. Jiang, T. Black, and
information acceptance and balance user beliefs. H. Liu, “A survey on echo chambers on social media: Description,
In summary, experimental results demonstrate that RGRec detection and mitigation,” 2021.
excels in mitigating filter bubble effects, underscoring its [6] C. Yu, L. Lakshmanan, and S. Amer-Yahia, “It takes variety to make
a world: diversification in recommender systems,” in Proceedings of
substantial contribution to the evolution of recommendation the 12th international conference on extending database technology:
systems. Beyond just enhancing recommendation diversity, Advances in database technology, 2009, pp. 368–378.
RGRec fosters belief harmony, carrying meaningful implica- [7] Q. Grossetti, C. Du Mouza, and N. Travers, “Community-based recom-
mendations on twitter: avoiding the filter bubble,” in Web Information
tions for the elevation of user satisfaction. These insights play Systems Engineering–WISE 2019, 2019, pp. 212–227.
a pivotal role in guiding the development and refinement of [8] G. M. Lunardi, G. M. Machado, V. Maran, and J. P. M. de Oliveira,
future personalized recommendation models and strategies. “A metric for filter bubble measurement in recommender algorithms
considering the news domain,” Applied Soft Computing, vol. 97, p.
106771, 2020.
VII. C ONCLUSION AND F UTURE W ORK [9] L. Yang, S. Wang, Y. Tao, J. Sun, X. Liu, P. S. Yu, and T. Wang, “Dgrec:
Graph neural network for recommendation with diversified embedding
In this paper, we focus on addressing the issue of filter generation,” in Proceedings of the Sixteenth ACM International Confer-
bubbles in recommendation systems and propose a Respon- ence on Web Search and Data Mining, 2023, pp. 661–669.
[10] Z. Li, Y. Dong, C. Gao, Y. Zhao, D. Li, J. Hao, K. Zhang, Y. Li, and
sible Graph-based Recommendation, RGRec, as a solution to Z. Wang, “Breaking filter bubble: A reinforcement learning framework
mitigate the negative effects of filter bubbles by promoting of controllable recommender system,” in Proceedings of the ACM Web
belief harmony among users. Conference 2023, 2023, pp. 4041–4049.
[11] Q. Guo, F. Zhuang, C. Qin, H. Zhu, X. Xie, H. Xiong, and Q. He,
RGRec is an intermediary between existing preference- “A survey on knowledge graph-based recommender systems,” IEEE
based recommendation systems and users, aiming to facilitate Transactions on Knowledge and Data Engineering, vol. 34, no. 8, pp.
democratic and transparent recommendations. It incorporates 3549–3568, 2020.
[12] M. Jesse and D. Jannach, “Digital nudging with recommender systems:
several key features. First, it utilizes FBDetect for filter bub- Survey and future directions,” Computers in Human Behavior Reports,
ble identification. Second, it applies nudging techniques to vol. 3, p. 100052, 2021.
broaden users’ interests and balance their beliefs incremen- [13] S. Joachim, A. R. M. Forkan, P. P. Jayaraman, A. Morshed, and
N. Wickramasinghe, “A nudge-inspired ai-driven health platform for
tally. Finally, it incorporates a user feedback loop based on self-management of diabetes,” Sensors, vol. 22, no. 12, p. 4620, 2022.
RecomGen to capture evolving user beliefs over time and [14] W. Li, Q. Bai, and M. Zhang, “Siminer: A stigmergy-based model for
increase recommendation diversity. mining influential nodes in dynamic social networks,” IEEE Transac-
tions on Big Data, vol. 5, no. 2, pp. 223–237, 2019.
The experimental results explicitly reveal the effectiveness [15] G. K. Patro, A. Biswas, N. Ganguly, K. P. Gummadi, and
of RGRec in mitigating filter bubbles and balancing user A. Chakraborty, “Fairrec: Two-sided fairness for personalized recom-
beliefs. For future research, we plan to explore additional mendations in two-sided platforms,” in Proceedings of the web confer-
ence 2020, 2020, pp. 1194–1204.
techniques to augment the model’s performance further, un- [16] Q. Guo, Z. Sun, J. Zhang, and Y.-L. Theng, “An attentional recurrent
dertake user studies to assess the long-term impacts, and neural network for personalized next location recommendation,” in
investigate the influence of RGRec on critical thinking abilities Proceedings of the AAAI Conference on artificial intelligence, vol. 34,
no. 01, 2020, pp. 83–90.
and inclusivity. Persistent refinement and evaluation of the [17] F. Wang, H. Zhu, G. Srivastava, S. Li, M. R. Khosravi, and L. Qi,
model could lead to improved recommendation systems that “Robust collaborative filtering recommendation with user-item-trust
cater more effectively to the diverse needs of users. records,” IEEE Transactions on Computational Social Systems, vol. 9,
no. 4, pp. 986–996, 2021.
[18] S. Reddy, S. Nalluri, S. Kunisetti, S. Ashok, and B. Venkatesh, “Content-
ACKNOWLEDGMENTS based movie recommendation system using genre correlation,” in Smart
Intelligent Computing and Applications: Proceedings of the Second
The authors would like to acknowledge the financial support International Conference on SCI 2018, Volume 2, 2019, pp. 391–397.
from Callaghan Innovation (CSITR1901, 2021), New Zealand, [19] Z. Wu, C. Li, J. Cao, and Y. Ge, “On scalability of association-rule-
without which this research would not have been possible. based recommendation: A unified distributed-computing framework,”
ACM Transactions on the Web (TWEB), vol. 14, no. 3, pp. 1–21, 2020.
We are grateful for their contributions to the advancement of [20] Y. Afoudi, M. Lazaar, and M. Al Achhab, “Hybrid recommendation
science and technology in New Zealand. The authors would system combined content-based filtering and collaborative prediction
also like to thank CAITO.ai for their invaluable partnership using artificial neural network,” Simulation Modelling Practice and
Theory, vol. 113, p. 102375, 2021.
and their contributions to the project. [21] L. V. Bryant, “The youtube algorithm and the alt-right filter bubble,”
Open Information Science, vol. 4, no. 1, pp. 85–90, 2020.
[22] B. Kitchens, S. L. Johnson, and P. Gray, “Understanding echo chambers
R EFERENCES and filter bubbles: The impact of social media on diversification and
[1] H. Ko, S. Lee, Y. Park, and A. Choi, “A survey of recommendation partisan shifts in news consumption.” MIS quarterly, vol. 44, no. 4,
systems: recommendation models, techniques, and application fields,” 2020.
Electronics, vol. 11, no. 1, p. 141, 2022. [23] P. M. Dahlgren, “A critical review of filter bubbles and a comparison
[2] L. Michiels, J. Leysen, A. Smets, and B. Goethals, “What are filter with selective exposure,” Nordicom Review, vol. 42, no. 1, pp. 15–33,
bubbles really? a review of the conceptual and empirical work,” in 2021.
Adjunct Proceedings of the 30th ACM Conference on User Modeling, [24] J. Chen, H. Dong, X. Wang, F. Feng, M. Wang, and X. He, “Bias and
Adaptation and Personalization, 2022, pp. 274–279. debias in recommender system: A survey and future directions,” ACM
Transactions on Information Systems, vol. 41, no. 3, pp. 1–39, 2023.

14 JOURNAL OF IEEE TRANSACTIONS ON ARTIFICIAL INTELLIGENCE, VOL. 00, NO. 0, NOV. 2023

[25] H. Tang, S. Wu, G. Xu, and Q. Li, “Dynamic graph evolution learning for [47] Y. Hao, L. Dong, F. Wei, and K. Xu, “Visualizing and understanding
recommendation,” in Proceedings of the 46th International ACM SIGIR the effectiveness of BERT,” in Proceedings of the 2019 Conference on
Conference on Research and Development in Information Retrieval, Empirical Methods in Natural Language Processing and the 9th Inter-
2023, pp. 1589–1598. national Joint Conference on Natural Language Processing (EMNLP-
[26] D. P. Calvillo, A. B. Swan, and A. M. Rutchick, “Ideological belief IJCNLP), 2019, pp. 4143–4152.
bias with political syllogisms,” Thinking & Reasoning, vol. 26, no. 2, [48] Q. Zhang, J. Li, Q. Jia, C. Wang, J. Zhu, Z. Wang, and X. He, “Unbert:
pp. 291–310, 2020. User-news matching bert for news recommendation,” in Proceedings of
[27] P. Resnick, R. K. Garrett, T. Kriplean, S. A. Munson, and N. J. Stroud, the Thirtieth International Joint Conference on Artificial Intelligence,
“Bursting your (filter) bubble: strategies for promoting diverse expo- 2021, pp. 3356–3362.
sure,” in Proceedings of the 2013 conference on Computer supported [49] C. Jeong, S. Jang, E. Park, and S. Choi, “A context-aware citation
cooperative work companion, 2013, pp. 95–100. recommendation model with bert and graph convolutional networks,”
[28] Y. Hu, S. Wu, C. Jiang, W. Li, Q. Bai, and E. Roehrer, “Ai facilitated Scientometrics, vol. 124, pp. 1907–1922, 2020.
isolations? the impact of recommendation-based influence diffusion in [50] Z. Qiu, X. Wu, J. Gao, and W. Fan, “U-bert: Pre-training user repre-
human society,” in Proceedings of the Thirty-First International Joint sentations for improved recommendation,” in Proceedings of the AAAI
Conference on Artificial Intelligence, 2022, pp. 5080–5086. Conference on Artificial Intelligence, vol. 35, no. 5, 2021, pp. 4320–
[29] W. Wang, F. Feng, L. Nie, and T.-S. Chua, “User-controllable recommen- 4327.
dation against filter bubbles,” in Proceedings of the 45th International [51] X. Wang, X. He, M. Wang, F. Feng, and T.-S. Chua, “Neural graph
ACM SIGIR Conference on Research and Development in Information collaborative filtering,” in Proceedings of the 42nd international ACM
Retrieval, 2022, pp. 1251–1261. SIGIR conference on Research and development in Information Re-
[30] U. Chitra and C. Musco, “Analyzing the impact of filter bubbles on trieval, 2019, pp. 165–174.
social network polarization,” in Proceedings of the 13th International [52] X. He, K. Deng, X. Wang, Y. Li, Y. Zhang, and M. Wang, “Lightgcn:
Conference on Web Search and Data Mining, 2020, pp. 115–123. Simplifying and powering graph convolution network for recommenda-
[31] F. Xiao, “Multi-sensor data fusion based on the belief divergence tion,” in Proceedings of the 43rd International ACM SIGIR conference
measure of evidences and the belief entropy,” Information Fusion, on research and development in Information Retrieval, 2020, pp. 639–
vol. 46, pp. 23–32, 2019. 648.
[32] J. Blumenthal-Barby and D. J. Opel, “Nudge or grudge? choice archi- [53] X. Wang, H. Jin, A. Zhang, X. He, T. Xu, and T.-S. Chua, “Disentangled
tecture and parental decision-making,” Hastings Center Report, vol. 48, graph collaborative filtering,” in Proceedings of the 43rd international
no. 2, pp. 33–39, 2018. ACM SIGIR conference on research and development in information
[33] J. Beshears and H. Kosowsky, “Nudging: Progress to date and future retrieval, 2020, pp. 1001–1010.
directions,” Organizational behavior and human decision processes, vol. [54] J. P. Shaver, “What statistical significance testing is, and what it is not,”
161, pp. 3–19, 2020. The Journal of Experimental Education, vol. 61, no. 4, pp. 293–316,
[34] S. Wu, W. Li, H. Shen, and Q. Bai, “Identifying influential users in 1993.
unknown social networks for adaptive incentive allocation under budget
restriction,” Information Sciences, vol. 624, pp. 128–146, 2023. Mengyan Wang received her master’s degree in
[35] S. Wu, W. Li, and Q. Bai, “Gac: A deep reinforcement learning model Computer and Information Sciences from Auckland
toward user incentivization in unknown social networks,” Knowledge- University of Technology, New Zealand, in 2019.
Based Systems, vol. 259, p. 110060, 2023. Currently, she is pursuing a Ph.D. degree at Auck-
[36] W. Li, Q. Bai, M. Zhang, and T. D. Nguyen, “Automated Influence land University of Technology, and her research
Maintenance in Social Networks: an Agent-based Approach,” IEEE mainly focuses on recommendation systems and
Transactions on Knowledge and Data Engineering, vol. 31, no. 10, pp. knowledge graphs.
1884–1897, 2019.
[37] G. Wang, W. Li, Q. Bai, and E. M.-K. Lai, “Maximizing social influence
with minimum information alteration,” IEEE Transactions on Emerging
Topics in Computing, pp. 1–13, 2023.
[38] S. Wu, Q. Bai, and S. Sengvong, “Greencommute: An influence- Yuxuan Hu received her Ph.D. from the University
aware persuasive recommendation approach for public-friendly commute of Tasmania, Australia, in 2023, the M.E degree
options,” Journal of Systems Science and Systems Engineering, vol. 27, from the University of Auckland, New Zealand, in
no. 2, pp. 250–264, 2018. 2017, and the bachelor’s degree from Hunan Normal
[39] D.-A. Sitar-Tăut, D. Mican, and R. A. Buchmann, “A knowledge- University, China, in 2016. Her research interests in-
driven digital nudging approach to recommender systems built on a clude agent-based modeling, social network analysis,
modified onicescu method,” Expert Systems with Applications, vol. 181, and recommendation systems.
p. 115170, 2021.
[40] E. Million, “The hadamard product,” Course Notes, vol. 3, no. 6, pp.
1–7, 2007.
[41] M. Behrendt and S. Harmeling, “Arguebert: How to improve bert
embeddings for measuring the similarity of arguments,” in Proceedings
of the 17th Conference on Natural Language Processing (KONVENS Shiqing Wu is a Postdoctoral Research Associate
2021), 2021, pp. 28–36. in the School of Computer Science at the University
[42] J. D. M.-W. C. Kenton and L. K. Toutanova, “Bert: Pre-training of deep of Technology Sydney. He received his Ph.D. from
bidirectional transformers for language understanding,” in Proceedings the University of Tasmania in 2022 and a joint
of the 2019 Conference of the North American Chapter of the Asso- B.Sc. from Auckland University of Technology and
ciation for Computational Linguistics: Human Language Technologies, China Jiliang University in 2016. His research inter-
vol. 1, 2019, p. 2. ests involve recommendation systems, reinforcement
[43] A. Wooditch, N. J. Johnson, R. Solymosi, J. Medina Ariza, and S. Lang- learning, social influence analysis, and agent-based
ton, “The normal distribution and single-sample significance tests,” A modeling. He has published several research papers
Beginner’s Guide to Statistics for Criminology and Criminal Justice in prestigious venues, such as SIGIR, ICDM, IJCAI,
Using R, pp. 155–168, 2021. AAMAS, etc.
[44] D. J. Steinskog, D. B. Tjøstheim, and N. G. Kvamstø, “A cautionary
note on the use of the kolmogorov–smirnov test for normality,” Monthly
Weather Review, vol. 135, no. 3, pp. 1151–1157, 2007.
[45] M. Barbehenn, “A note on the complexity of dijkstra’s algorithm for
graphs with weighted vertices,” IEEE transactions on computers, vol. 47,
no. 2, p. 263, 1998.
[46] G. Ramalingam and T. Reps, “A categorized bibliography on incremental
computation,” in Proceedings of the 20th ACM SIGPLAN-SIGACT
symposium on Principles of programming languages, 1993, pp. 502–
510.

WANG et al.: NUDGING TOWARDS RESPONSIBLE RECOMMENDATIONS 15

Weihua Li received his Ph.D. from Auckland Uni-

versity of Technology, New Zealand in 2018 and his
M.Tech from the National University of Singapore in
2014. He is currently a senior lecturer in the School
of Engineering, Computer & Mathematical Sciences.
Weihua is an active researcher, and his research
interest mainly focuses on Artificial Intelligence,
including agent-based modeling and simulation in
complex systems and natural language processing.
He has over 40 publications and is involved in the or-
ganization of a number of international conferences
in the field of AI.

Quan Bai is a distinguished researcher with a

primary focus on machine learning, agent-based
modeling and knowledge representation. He earned
his Ph.D. in Computer Science from the University
of Wollongong in 2007. Over the years, Dr. Bai’s
career has encompassed a range of pivotal roles
in the field. He began as a postdoctoral research
fellow at the Commonwealth Scientific and Indus-
trial Research Organisation (CSIRO) in Australia.
Subsequently, he expanded his horizons as a lecturer
and later as a senior lecturer at Auckland University
of Technology in New Zealand. Presently, Dr. Bai serves as an associate
professor at the University of Tasmania and leads an active AI Research
Group at the university. Dr. Bai has published over 160 research publications,
underscoring his commitment to advancing in his field.

Zihan Yuan received her Bachelor’s degrees in

2019 and Master’s degree in 2022 from Monash
University, Australia. She is currently a Master of
Information Technology and Systems student at the
University of Tasmania, pursuing the research path-
way in responsible recommendation systems.

Chenting Jiang holds a Bachelor’s degree in Ge-

ography and Economics (BE in Geography and BEc
in Economics) from South China Normal University,
earned in 2011. In 2021, she achieved a Master’s
degree in Information and Communication Technol-
ogy from the University of Tasmania. Her research
focuses on diverse areas, including machine learn-
ing, neural networks, data assimilation, predicting
soil water retention function, and hybrid hydraulic
modeling. Additionally, she explores topics in online
social networks, influence diffusion, and behavioral
economics.

Responsible Tech Guide by All Tech Is Human
93% (30)
Responsible Tech Guide by All Tech Is Human
160 pages
1 s2.0 S1566253523002221 Main
No ratings yet
1 s2.0 S1566253523002221 Main
22 pages
Filter Bubbles in Recommender Systems Fact
No ratings yet
Filter Bubbles in Recommender Systems Fact
21 pages
A Graph-Based Approach For Mitigating Multi-Sided Exposure Bias
No ratings yet
A Graph-Based Approach For Mitigating Multi-Sided Exposure Bias
31 pages
Ds23f Safa Ali Khan m23f0022ds030 Assg#1
No ratings yet
Ds23f Safa Ali Khan m23f0022ds030 Assg#1
2 pages
A Survey On Reinforcement Learning For Recommender Systems
No ratings yet
A Survey On Reinforcement Learning For Recommender Systems
21 pages
Modeling Recommender Ecosystems: Research Challenges at The Intersection of Mechanism Design, Reinforcement Learning and Generative Models
No ratings yet
Modeling Recommender Ecosystems: Research Challenges at The Intersection of Mechanism Design, Reinforcement Learning and Generative Models
24 pages
On Deep Learning For Trust-Aware Recommendations in Social Networks
No ratings yet
On Deep Learning For Trust-Aware Recommendations in Social Networks
14 pages
多样化嵌入的图神经网络3570472
No ratings yet
多样化嵌入的图神经网络3570472
9 pages
Appendix
No ratings yet
Appendix
9 pages
Mirza 2003
No ratings yet
Mirza 2003
30 pages
Web Crawling Based Context Aware Recommender Syste
No ratings yet
Web Crawling Based Context Aware Recommender Syste
25 pages
Recommender Systems and Their Ethical Challenges
No ratings yet
Recommender Systems and Their Ethical Challenges
12 pages
Bitirme
No ratings yet
Bitirme
9 pages
03-Knowledge-Based Recommender Systems - Overview and Research Directions
No ratings yet
03-Knowledge-Based Recommender Systems - Overview and Research Directions
19 pages
Li 等 - 2023 - Fairness in Recommendation Foundations, Methods a
No ratings yet
Li 等 - 2023 - Fairness in Recommendation Foundations, Methods a
49 pages
6-Recommender Systems For Good RS4Good Survey of Use
No ratings yet
6-Recommender Systems For Good RS4Good Survey of Use
17 pages
Eecs 2024 159
No ratings yet
Eecs 2024 159
136 pages
Bias Resistant Retrieval Augmented Generation: A Clustering and BiQ Driven Approach With Equitable AI
No ratings yet
Bias Resistant Retrieval Augmented Generation: A Clustering and BiQ Driven Approach With Equitable AI
11 pages
Digital Nudging
No ratings yet
Digital Nudging
14 pages
An Effective Recommender System Based On
No ratings yet
An Effective Recommender System Based On
17 pages
No - Ntnu Inspera 74730513 28709119
No ratings yet
No - Ntnu Inspera 74730513 28709119
79 pages
Research Paper
No ratings yet
Research Paper
16 pages
T M R: A: Rust Etrics in Ecommender Systems Survey
No ratings yet
T M R: A: Rust Etrics in Ecommender Systems Survey
12 pages
Efficient Rec Springer
No ratings yet
Efficient Rec Springer
16 pages
Bayesian Probabilistic Matrix Factorization With Social Relations and Item Contents
No ratings yet
Bayesian Probabilistic Matrix Factorization With Social Relations and Item Contents
14 pages
ISMIS 2020 Tduricic
No ratings yet
ISMIS 2020 Tduricic
60 pages
Responsible AI White Paper by GG
No ratings yet
Responsible AI White Paper by GG
4 pages
Bauer 2024 Dagstuhl Perspectives
100% (2)
Bauer 2024 Dagstuhl Perspectives
116 pages
An Overview of Recommender Systems in Requirements Engineering (Preprint Version)
No ratings yet
An Overview of Recommender Systems in Requirements Engineering (Preprint Version)
18 pages
Recent Developments in Recommender Systems: A Survey
No ratings yet
Recent Developments in Recommender Systems: A Survey
21 pages
GS-RS A Generative Approach For Alleviating Cold Start and Filter Bubbles in Recommender Systems
No ratings yet
GS-RS A Generative Approach For Alleviating Cold Start and Filter Bubbles in Recommender Systems
14 pages
1 s2.0 S0306457321001369 Main
No ratings yet
1 s2.0 S0306457321001369 Main
18 pages
Slide Data Mining
No ratings yet
Slide Data Mining
46 pages
Using Graph Neural Networks For Social Recommendat
No ratings yet
Using Graph Neural Networks For Social Recommendat
18 pages
TKDE2022 提到了中间负样本这个概念和我的有点像
No ratings yet
TKDE2022 提到了中间负样本这个概念和我的有点像
14 pages
New Laws and Regulation: Opportunities For BISE Research
No ratings yet
New Laws and Regulation: Opportunities For BISE Research
14 pages
A Trust and Semantic Based Approach For Social Recommendation
No ratings yet
A Trust and Semantic Based Approach For Social Recommendation
15 pages
KEOD 2024 37 Formatted
No ratings yet
KEOD 2024 37 Formatted
12 pages
Hybrid News Recommendation System Using TF-IDF and Machine Learning Approach
No ratings yet
Hybrid News Recommendation System Using TF-IDF and Machine Learning Approach
5 pages
How Recommendation Accuracy Induces Users Intermittent Discontin
No ratings yet
How Recommendation Accuracy Induces Users Intermittent Discontin
18 pages
RAI Guidelines: Method For Generating Responsible AI Guidelines Grounded in Regulations and Usable by (Non-) Technical Roles
No ratings yet
RAI Guidelines: Method For Generating Responsible AI Guidelines Grounded in Regulations and Usable by (Non-) Technical Roles
28 pages
The Behavioral Code Recommender Systems
No ratings yet
The Behavioral Code Recommender Systems
17 pages
Research Article: An Effective Recommender Algorithm For Cold-Start Problem in Academic Social Networks
No ratings yet
Research Article: An Effective Recommender Algorithm For Cold-Start Problem in Academic Social Networks
12 pages
Trustworthy GNNs New Version
No ratings yet
Trustworthy GNNs New Version
42 pages
Irjet V6i7379
No ratings yet
Irjet V6i7379
9 pages
19.SocialLGN Light Graph Convolution Network For Social Recommendation
No ratings yet
19.SocialLGN Light Graph Convolution Network For Social Recommendation
13 pages
Anti-Fakeu: Defending Shilling Atacks On Graph Neural Network Based Recommender Model
No ratings yet
Anti-Fakeu: Defending Shilling Atacks On Graph Neural Network Based Recommender Model
11 pages
Bursting Filter Bubbles by Counterfactual Interactive Recommender System
No ratings yet
Bursting Filter Bubbles by Counterfactual Interactive Recommender System
26 pages
Slide Data Mining
No ratings yet
Slide Data Mining
46 pages
A Troubling Analysis of Reproducibility and Progress in Recommender Systems Research
No ratings yet
A Troubling Analysis of Reproducibility and Progress in Recommender Systems Research
49 pages
Recommender Systems: A, B, C B B, D A, B, C, A, B, C A, B, C, e
No ratings yet
Recommender Systems: A, B, C B B, D A, B, C, A, B, C A, B, C, e
97 pages
Bauer 2024 Values
No ratings yet
Bauer 2024 Values
40 pages
RS Unit - 3,4
No ratings yet
RS Unit - 3,4
4 pages
Responsible Innovation 2 PDF
No ratings yet
Responsible Innovation 2 PDF
312 pages
Users Choose To Engage With More Partisan News Than They Are Exposed To On Google Search
No ratings yet
Users Choose To Engage With More Partisan News Than They Are Exposed To On Google Search
25 pages
RS PDF
No ratings yet
RS PDF
5 pages
Environmental Mode Choice Nudge
No ratings yet
Environmental Mode Choice Nudge
13 pages
Commenting On Given Cases
No ratings yet
Commenting On Given Cases
10 pages
The Circuit
No ratings yet
The Circuit
4 pages
Cics Refresher
100% (1)
Cics Refresher
32 pages
Insert Grade
No ratings yet
Insert Grade
16 pages
Hts Log
No ratings yet
Hts Log
721 pages
Preparation of Standard Solutions
No ratings yet
Preparation of Standard Solutions
2 pages
Hotel Management Project Proposel
No ratings yet
Hotel Management Project Proposel
3 pages
Relationship Between International and Municipal Law
No ratings yet
Relationship Between International and Municipal Law
6 pages
Song List: No Penyanyi Judul Lagu
No ratings yet
Song List: No Penyanyi Judul Lagu
23 pages
FMEA - A Guide For Continuous Improvement
100% (1)
FMEA - A Guide For Continuous Improvement
36 pages
@canotes - Final Ca-Final-Sfm-New-Icai-Question-Paper-Nov-2020
No ratings yet
@canotes - Final Ca-Final-Sfm-New-Icai-Question-Paper-Nov-2020
11 pages
RTA-33.1 Crank Pin Bearing Shells
No ratings yet
RTA-33.1 Crank Pin Bearing Shells
13 pages
Book Project Model
No ratings yet
Book Project Model
8 pages
Peusangan 1 & 2 Hepp Construction Project
No ratings yet
Peusangan 1 & 2 Hepp Construction Project
4 pages
ASUS Back To College Offer: Wed, Jun 15, 2022 at 9:47 PM
No ratings yet
ASUS Back To College Offer: Wed, Jun 15, 2022 at 9:47 PM
1 page
Read The Following Advertisement and Mark The Letter 2
No ratings yet
Read The Following Advertisement and Mark The Letter 2
11 pages
Kota Kinabalu
No ratings yet
Kota Kinabalu
36 pages
Tazewell County Drug Charges
No ratings yet
Tazewell County Drug Charges
11 pages
Music Production Terms
No ratings yet
Music Production Terms
38 pages
IRPSM
No ratings yet
IRPSM
2 pages
1 Decision Making Problem Solving and Critical Thinking S
No ratings yet
1 Decision Making Problem Solving and Critical Thinking S
48 pages
The Origins of You How Breaking Family Patterns Can Liberate The Way We Live and Love Vienna Pharaon Instant Download
100% (1)
The Origins of You How Breaking Family Patterns Can Liberate The Way We Live and Love Vienna Pharaon Instant Download
39 pages
Chat GPT
No ratings yet
Chat GPT
28 pages
The On and Off Connection Between Political and Economic Progress
No ratings yet
The On and Off Connection Between Political and Economic Progress
7 pages
Rubric For Wind Vane
No ratings yet
Rubric For Wind Vane
1 page
Industrial IoT and Predictive Maintenance
No ratings yet
Industrial IoT and Predictive Maintenance
10 pages
9 Chap 9 Foundation of Group Behaviour
No ratings yet
9 Chap 9 Foundation of Group Behaviour
23 pages
JCAD February 2025
No ratings yet
JCAD February 2025
79 pages
Tri An-Prc 152
No ratings yet
Tri An-Prc 152
16 pages
Fatty Acid Composition of The Lipids of Membranes of Gram-Positive Bacteria and ,,walls" of Gram-Negative Bacteria
No ratings yet
Fatty Acid Composition of The Lipids of Membranes of Gram-Positive Bacteria and ,,walls" of Gram-Negative Bacteria
3 pages
Acceleration reportJM
No ratings yet
Acceleration reportJM
4 pages