Recommender Systems: An Overview

Robin Burke1, Alexander Felfernig2, Mehmet H. Göker3

1
Web Intelligence Laboratory, DePaul University, [email protected]
2
Graz University of Technology Graz, Austria, [email protected]
3
Salesforce.com, The Landmark @ One Market, San Francisco, CA 94105, [email protected]

Abstract provides a personalized view of the data, in this case, the

Recommender systems are tools for interacting with large bookstore’s inventory. If we take away the personalization,
and complex information spaces. They provide a we are left with the list of best-sellers – a list that is
personalized view of such spaces, prioritizing items likely to independent of the user. The aim of the recommender
be of interest to the user. The field, christened in 1995, has
system is to lower the user’s search effort by listing those
grown enormously in the variety of problems addressed and
techniques employed, as well as in its practical applications. items of highest utility, those that Jane might be most
Recommender systems research has incorporated a wide likely to purchase. This, of course, is beneficial to Jane as
variety of artificial intelligence techniques including well as the e-commerce store owner.
machine learning, data mining, user modeling, case-based Recommender systems research encompasses scenarios
reasoning, and constraint satisfaction, among others.
like this and many other information access environments
Personalized recommendations are an important part of
many on-line e-commerce applications such as in which a user and store owner can benefit from the
Amazon.com, Netflix, and Pandora. This wealth of practical presentation of personalized options. The field has seen a
application experience has provided inspiration to tremendous expansion of interest in the past decade,
researchers to extend the reach of recommender systems catalyzed in part by the Netflix Prize (Bennett & Lanning,
into new and challenging areas. The purpose of this special
2007) and evidenced by the rapid growth of the annual
issue is to take stock of the current landscape of
recommender systems research and identify directions the ACM Recommender Systems conference. At this point, it
field is now taking. This article provides an overview of the is worthwhile to take stock, to consider what distinguishes
current state of the field and introduces the various articles recommender systems research from other related areas of
in the special issue. research in artificial intelligence, and to examine the field’s
successes and new challenges.
Introduction
The prototypical use case for a recommender system What is a Recommender System?
occurs regularly in e-commerce settings. A user, Jane, The definition of a recommender system has evolved over
visits her favorite online bookstore. The homepage lists the past 14 years. In Resnick and Varian’s seminal article,
current bestsellers and also a list containing recommended the authors describe a recommender system as follows:
items. This list might include, for example, a new book
published by one of Jane’s favorite authors, a cookbook by
a new author and a supernatural thriller. Whether Jane will “In a typical recommender system people
find these suggestions useful or distracting is a function of provide recommendations as inputs, which the
system then aggregates and directs to
how well they match her tastes. Is the cookbook for a style
appropriate recipients. In some cases the
of cuisine that she likes (and is it different enough from
primary transformation is in the aggregation;
ones she already owns)? Is the thriller too violent? A key
in others the system’s value lies in its ability
feature of a recommender system therefore is that it to make good matches between the
recommenders and those seeking
Copyright © 2011, Association for the Advancement of Artificial recommendations.” (Resnick & Varian, 1997)
Intelligence (www.aaai.org). All rights reserved.
 Note that this definition places the emphasis on the results relevant to a particular query to be the same
recommender systems as supporting the collaboration regardless of who issued it.1 Many recommender systems
between users. Later researchers have expanded the achieve personalization by maintaining profiles of user’s
definition to include systems that suggest items of interest, activity (long-term or short-term) or stated preferences
regardless of how those recommendations are produced: (Schafer, et al. 2007). Others achieve a personalized result
“any system that produces individualized through conversational interaction (McGinty & Reilly,
recommendations as output or has the effect of guiding the 2011).
user in a personalized way to interesting or useful objects
in a large space of possible options.” (Burke, 2002) This
more general definition was formalized by Adomavicius A Recommender System Typology
and Tuzhilin (Adomavicius & Tuzhilin, 2005): Recommender systems research is characterized by a
More formally, the recommendation problem common problem area rather than a common technology or
can be formulated as follows: Let C be the set approach. An examination of the past four ACM
of all users and let S be the set of all possible Recommender System conferences shows that a wide
items that can be recommended...Let u be a variety of research approaches have been applied to the
utility function that measures the usefulness of recommender systems problem, from statistical methods to
item s to user c, i.e., u: CxSŸR, where R is a ontological reasoning, and a wide variety of problems have
totally ordered set (e.g., nonnegative integers been tackled, from choosing consumer products to finding
or real numbers within a certain range). Then, friends and lovers.
for each user c‫א‬C, we want to choose such One lesson that has been learned over the past years of
item s'‫א‬S that maximizes the user’s utility. recommender systems research is that the application
domain exerts a strong influence over the types of methods
This definition opens up the field of recommender that can be successfully applied. Domain characteristics
systems to any application that computes a user-specific like the persistence of the user’s utility function have a big
utility, encompassing many problems commonly thought impact: for example, a users’ taste in music may change
of as database or information retrieval applications. Even slowly but his interest in celebrity news stories may
this broad definition may be too narrow as some fluctuate much more. Thus, the reliability of preferences
recommenders may operate on configurations – as opposed gathered in the past may vary. Similarly, some items, such
to a fixed set S of all items – and others make as books, are available for recommendation and
recommendations for groups (utility is computed for a consumption over a long period of time – often years. On
subset C*‫ؿ‬C of the users rather than a single user.) The the other hand, in a technological domain, such as cell
definition may also be a bit misleading in that many phones or cameras, old products become rapidly obsolete
recommender systems do not explicitly calculate utilities and cannot be usefully recommended. This is also true of
when they produce a ranked list of recommended items. areas where timeliness matters such as news and cultural
The authors are careful to say that the goal is to choose the events. See (Burke & Ramezani, 2011) for a more
items with the best utility, not necessarily to compute the complete description of the factors that influence the
utility in some explicit way. choice of recommendation approach.
From these considerations, two basic principles stand It is not surprising therefore that there are multiple
out that distinguish recommender systems research strands of research in recommender systems, as researchers
tackle a variety of recommendation domains. To unify
• A recommender system is personalized. The these disparate approaches, it is useful to consider the AI
recommendations it produces are meant to aspects of recommendation, in particular, the knowledge
optimize the experience of one user, not to basis underlying a recommender system.
represent group consensus for all.
Knowledge Sources
• A recommender system is intended to help the Every AI system draws on one or more sources of
user select among discrete options. Generally knowledge in order to do its work. A supervised machine
the items are already known in advance and learning system, for example, would have a labeled
not generated in a bespoke fashion.
collection of data as its primary knowledge source, but the
The personalization aspect of recommender systems
algorithm and its parameters can be considered another
distinguishes this line of research most strongly from what
is commonly understood as research in search engines and
other information retrieval applications. In a search engine
or other information retrieval system, we expect the set of 1
Personalized search, for example, removes this distinction, and is
therefore by this definition a recommender systems application.
implicit kind of knowledge that is brought to bear on the be useful but are insufficient to make high-quality
classification task. recommendations in this domain.
Recommendation algorithms can also be classified
according to the knowledge sources that they use. Figure 1
shows the taxonomy of knowledge sources used in Research Questions in Recommender Systems
(Felfernig & Burke, 2008). There are three basic types of Collaborative Recommendation
knowledge: The most prominent technique in recommendation is
x social knowledge about the user base in general, collaborative recommendation. (Schafer et al., 2007) The
x individual knowledge about the particular user basic insight for this technique is a sort of continuity in the
for whom recommendations are sought (and realm of taste – if users Alice and Bob have the same
possibly knowledge about the specific utility for items 1 through k, then the chances are good that
requirements those recommendations need to they will have the same utility for item k+1. Usually, these
meet), and finally utilities are based on ratings that users have supplied for
x content knowledge about the items being items with which they are already familiar.
recommended, ranging from simple feature lists The key advantage of collaborative recommendation is
to more complex ontological knowledge and its simplicity. The problem of computing utility is
means-ends knowledge that enable the system transformed into the problem of extrapolating missing
to reason about how an item can meet a user’s values in the ratings matrix, the sparse matrix where each
needs. user is a row, each item a column, and the values are the
known ratings.
This insight can be operationalized in a number of ways.
Originally, nearest-neighbor techniques were applied to
find neighborhoods of like-minded peers. However, matrix
factorization and other dimensionality-reduction
techniques are now recognized as superior in
accuracy (Bell & Koren, 2007).
Some problems with collaborative recommendation are
well-established:
• New items cannot be recommended without
relying on some additional knowledge source.
Extrapolation depends on having some values
from which to project. Indeed, sparsely-rated
items in general present a problem because
the system lacks information on which to base
Figure 1: Taxonomy of knowledge sources in recommendation predictions. By the same token, users who
(after [Felfernig & Burke, 2008]). have supplied few ratings will receive noisier
Different recommendation approaches draw from recommendations than those with more
different parts of this spectrum of knowledge sources. The substantial histories. The problems of new
terms of the Netflix Prize competition made available only users and new ratings are collectively known
opinions in the form of ratings, but no requirements or as the “cold start” problem in collaborative
demographic information about users (Bennet & Lanning, recommendation.
2007). Good domain knowledge is notoriously difficult to • The distribution of ratings and user
assemble in this domain because of the complexity of preferences in many consumer taste domains
representing and reasoning about narrative content, is fairly concentrated: a small number of
directorial style, etc. The problem thus lent itself to a “blockbuster” items receive a great deal of
mathematical approximation technique working attention, and there are many, many rarely-
exclusively from ratings both social and individual (Bell, rated items.
Koren & Volinsky, 2007). • Malicious users may be able to generate large
By contrast, the problem of recommending investment numbers of pseudonymous “sybil” profiles
options reported in (Felfernig & Burke, 2008) can benefit and use them to bias the recommendations of
from detailed knowledge about the customer’s income and the system in one way or another.
financial status, the other items in their portfolio, and their
attitude toward risk. Other users’ opinions and choices may
 There is still a great deal of algorithmic research focused and how these interact with the financial position of the
on the problems of collaborative recommendation: more investor.
accurate and efficient estimates of the ratings matrix, better As with other knowledge-based systems, knowledge
handling of new users and new items, and the extension of acquisition, maintenance and validation are key issues.
the basic collaborative recommendation idea to new types Also, since knowledge-based recommenders can make use
of data including multi-dimensional ratings and user- of detailed requirements from the user, user interface
generated tags, among others. research has been paramount in developing knowledge-
Content-based Recommendation based recommenders that do not place too much of a
Before the advent of collaborative recommendation in the burden on users.
1990s, earlier research in personalized information access Evaluation
had concentrated on combining knowledge about items Because of the difficulties of running large-scale user
with information about user’s preferences in order to locate studies, recommender systems have conventionally been
appropriate items. Both Rich;s early work on Grundy evaluated on one or both of the following measures:
(book recommendation) (Rich, 1979) and Rocchio’s • Prediction accuracy. How well do the system’s
method (information retrieval) (Rocchio, 1971) can now be predicted ratings compare with those that are
seen as early examples of recommender systems although known, but withheld?
the term had not yet been coined. This approach, because
• Precision of recommendation lists. Given a short
of its reliance on the content knowledge source, in
list of recommendations produced by the system
particular, item features, has come to be known as content-
(typically all a user would have patience to
based recommendation. examine), how many of the entries match known
Content-based recommendation is closely linked with “liked” items?
supervised machine learning. We can view the problem as
one of learning a set of user-specific classifiers where the Both of these conventional measures are deficient in
classes are “useful to user X” and “not useful to user X”. some key respects and many of the new areas of
One of the key issues in content-based recommendation exploration in recommender systems have led to
is feature quality. The objects to be recommended need to experimentation with new evaluation metrics to
be described so that meaningful learning of user supplement these common ones.
preferences can occur. Ideally, every object would be One of the most significant problems occurs because of
described at the same level of detail and the feature set the long-tailed nature of the ratings distribution in many
would contain descriptors that correlate with the datasets. A recommendation technique that optimizes for
discriminations made by users. Unfortunately, this is often high accuracy over the entire data set therefore contains a
not the case. Descriptions may be partial or some parts of implicit bias towards well-known items, and therefore may
the object space may be described in greater detail than fail to capture aspects of utility related to novelty. An
others. accurate prediction on an item that the user already knows
The match between the feature set and the user’s utility is inherently less useful than a prediction on an obscure
function also needs to be good. One of the strengths of the item.
popular Pandora streaming music service is that the To address this issue, some researchers are looking at
feature set it uses for musical selections are manually the balance between accuracy and diversity in a set of
chosen by music-savvy listeners. Automatic music recommendations, and working on algorithms that are
processing is not yet good enough to reliably extract sensitive to item distributions.
features like “bop feel” from a Charlie Parker recording. Another problem with conventional recommender
In addition to the development and application of new systems evaluation is that it is essentially static. A fixed
learning algorithms for the recommendation task, research database of ratings is divided into training and test sets and
in content-based recommendation also examines the used to demonstrate the effectiveness of an algorithm.
problem of feature extraction in different domains. However, the user experience of recommendation is quite
A further subtype of content-based recommendation is different. In an application like movie recommendation,
knowledge-based recommendation, in which the reliance the field of items is always expanding; a user’s tastes are
on item features is extended to other kinds of knowledge evolving; new users are coming to the system. Some
about products and their potential utilities for users. An recommendation applications require that we take the
example of this kind of system is the investment dynamic nature of the recommendation environment into
recommender mentioned earlier that has to know about the account, and evaluate our algorithms accordingly.
risk profiles and tax consequences of different investments Another area of evaluation that is relatively
underexamined is the interaction between the utility
functions of the store owner and the user, which approaches to music recommendation, explain these
necessarily look quite different. Owners implement recommendation styles using data from last.fm, and point
recommender systems in order to achieve business goals, out research challenges that need to be addressed.
typically increased profit. The owner therefore may prefer Adomavicius et al. analyze the impact of context on
an imperfect match with a high profit margin to a perfect recommender systems. Context aware recommendation
match with limited profit. On the other hand, a user who is goes beyond what we normally would consider
presented with low utility recommendations may cease to personalization and takes additional information such as
trust the recommendation function or the entire site. the environment and conditions the user is operating under
Owners with high volume sites can field algorithms side- into account. The authors analyze how these contextual
by-side in randomized trials and observe sales and profit factors can change and how they impact system design.
differentials, but such results rarely filter out into the They provide examples of implementations and describe
research community. challenges and future research directions for this field.
Software engineering is a relatively new application area
for recommender systems. Mobasher and Cleland-Huang
This Special Issue show how various areas of requirements engineering can
The aim of this special issue is to give a brief overview of benefit from recommendation: stakeholder identification,
the history and current status of recommender system domain analysis, requirements elicitation and decision
research, to describe the current state of recommender support. They describe various approaches from literature
systems in practical use, and to highlight new directions in and point out areas for future research.
recommender systems research that may be of interest to Friedrich and Zanker look at the problem of explaining the
AI Magazine readers. The papers have been chosen to recommendations that a system gives. Such explanations
illuminate the state of the art in recommendation and to may be important in securing user confidence and
illustrate some of the challenges that must be faced in acceptance of recommended items. This paper describes
extending current techniques in recommendation to meet the various means by which recommender systems
new domains and new requirements. generate such explanations and points out open research
The special issue starts off with an article by Martin, et al. issues.
The authors, all of whom have considerable experience in The last paper in the issue by Falkner, Felfernig and Haag
both recommender systems research and industrial system looks at recommendation in domains with configurable
development, give a historical overview of the field, and products. Many of the techniques appropriate for pre-
describe their view of what the future holds for defined product catalogs are not appropriate for products
recommender systems research. The second article is by with many configurable parts as the number of possible
Susan Aldrich, an analyst from the Patricia Seybold group. complete configurations is exponential. The authors
This article looks at the commercial recommender systems discuss existing approaches to such configuration problems
landscape from the perspective of an industry analyst and and open issues.
describes how commercial recommender systems are We hope that the articles in the issue provide background
designed, deployed and evaluated. Given the maturity of information for researchers that are new to the field,
the field in terms of commercial applications, we found it guidance for researchers who want to commercialize their
important to convey the industrial viewpoint here. work, and new ideas and motivation to researchers who
Next, Smyth, et al. explain their work on collaborative web want to expand the already impressive amount of work in
search in the HeyStacks system (www.heystacks.com). the relatively young field of Recommender Systems.
The authors describe how the standard, one-size-fits-all As the editors of this issue, we would like to thank all of
web search can be made more personalized by using our authors and the AI Magazine for their hard work and
information about searches done by peers or collaborators. support.
Enhancing search with this social aspect increases the
quality of the results and makes them more relevant.
References
The theme of social recommendation continues with the
paper of Burke and his co-authors which highlights how G. Adomavicius and A. Tuzhilin, Toward the Next Generation of
recommender applications can leverage data coming from Recommender Systems: A Survey of the State-of-the-Art and
social applications and how recommendation algorithms Possible Extensions, IEEE Transactions on Knowledge and Data
need to advance to meet the new challenges these systems Engineering, pp. 734-749, June, 2005.
pose. R. M. Bell and Y. Koren, 2007 Scalable Collaborative Filtering
Celma and Lamere describe recommender systems in the with Jointly Derived Neighborhood Interpolation Weights,
field of music recommendation. They analyze existing
Proceedings of the 2007 Seventh IEEE International Conference
on Data Mining, p.43-52, October 28-31, 2007.
R. M. Bell, Y. Koren, and C. Volinsky. 2007 The BellKor
solution to the Net Flix Prize. Technical report, AT&T Labs
Research, 2007. http://www.netflixprize.com/assets/
ProgressPrize2007_KorBell.pdf.
J. Bennett and S. Lanning. 2007. The Net Flix Prize. In Proc. of
KDD Cup Workshop at SIGKDD'07, 13th ACM Int. Conf. on
Knowledge Discovery and Data Mining, pages 3-6, San Jose, CA,
USA, 2007.
R. Burke. 2002 Hybrid Recommender Systems: Survey and
Experiments. User Modeling and User-Adapted Interaction. 12,
4, 331-370.
R. Burke and M. Ramezani. 2011. Matching Recommendation
Technologies and Domains. In Ricci, Rokach, Shapira and Kantor
(eds.) Recommender Systems Handbook, pages 367-386.
Springer. 2011.
A. Felfernig and R. Burke. 2008. Constraint-based recommender
systems: technologies and research issues. In Proceedings of the
10th International Conference on Electronic Commerce (ICEC
'08). ACM, New York, NY. Article 3, 10 pages. 2008.
L. McGinty and J. Reilly. 2011. On the Evolution of Critiquing
Recommenders. In Ricci, Rokach, Shapira and Kantor (eds.)
Recommender Systems Handbook, pages 419-453. Springer.
2011.
P. Resnick and H. R. Varian. 1997. Recommender systems.
Commun. ACM 40, 3, 56-58. 1997
E. Rich. 1979. User Modeling via Stereotypes. Cognitive Science,
3, 4, 329-354. 1979
J. J. Rocchio. 1971. Relevance feedback in information retrieval.
In Salton (ed.) The SMART Retrieval System: Experiments in
Automatic Document Processing, pages 313-323. Prentice-Hall,
1971.
J. Schafer, D. Frankowski, J. Herlocker, and S. Sen. 2007.
Collaborative Filtering Recommender Systems. In Brusilovsky,
Kobsa, and Nejdl (eds.) The Adaptive Web, pages 291-324.
Lecture Notes in Computer Science 4321. Springer. 2007.