Recommender Systems in E-commerce

Abdul Rehman 200196

Qamar ul haq junaid 190377
May 26, 2024

Abstract
Recommender systems have evolved from being novelties on a few e-commerce sites to becoming critical business
tools that shape the landscape of online commerce. Many leading e-commerce platforms now employ recommender
systems to assist customers in discovering products they are likely to value. These systems learn from customer
interactions and suggest products tailored to individual preferences. This paper examines the role of recommender
systems in enhancing e-commerce sales and investigates six sites utilizing these systems, including some that employ
multiple recommendation algorithms. We propose a taxonomy of recommender systems based on these examples,
detailing the customer interfaces, underlying technologies, and required customer inputs. Additionally, we address the
significant energy consumption and environmental impact associated with these AI-driven systems. We emphasize the
need for energy-efficient designs that align with the United Nations Sustainable Development Goals (UNSDGs) while
maintaining high performance and accuracy. The paper concludes with suggestions for new, environmentally-conscious
applications of recommender systems in e-commerce.

Keywords
Electronic commerce, recommender systems, energy efficiency, customer engagement, sustainability, artificial intelli-
gence, UNSDGs.

1 INTRODUCTION
In his book *Mass Customization* (Pine, 1993), Joe Pine argues that companies need to shift from the old world of
mass production, characterized by standardized products, homogeneous markets, and long product life cycles, to a
new world where variety and customization are essential. Pine asserts that producing a single standardized product is
no longer sufficient; companies must develop multiple products to meet the diverse needs of multiple customers. The
rise of e-commerce has enabled businesses to offer customers more options, but this increased level of customization
also increases the amount of information customers must process to find what they need. Recommender systems offer
a solution to this information overload problem.
Recommender systems are utilized by e-commerce sites to suggest products to their customers. Recommendations
can be based on top overall sellers, customer demographics, or an analysis of past buying behavior to predict future pur-
chases. These techniques fall under the broader category of personalization, helping websites adapt to each customer.
Recommender systems automate personalization on the web, realizing Pine’s vision of individual customization. This
level of personalization is essential for modern e-commerce sites, as articulated by Jeff Bezos, CEO of Amazon.com™,
who stated, ”If I have 2 million customers on the Web, I should have 2 million stores on the Web.”
Recommender systems enhance e-commerce sales in three key ways:

1.1 Browsers into buyers:

Visitors often browse websites without making purchases. Recommender systems can guide customers to products
they are likely to buy, increasing conversion rates.

1.2 Cross-sell:
Recommender systems improve cross-selling by suggesting additional products. Effective recommendations can in-
crease the average order size by proposing complementary items during the checkout process.

1
1.3 Loyalty:
In a competitive online environment, customer loyalty is crucial. Recommender systems enhance loyalty by creating a
personalized shopping experience. Sites that invest in learning about their users and provide tailored recommendations
foster stronger customer relationships. Customers are more likely to return to sites that best meet their needs,
reinforcing their loyalty. As Pine et al. (1995) noted, once a customer has trained a recommendation system to
understand their preferences, switching to a competitor would require significant effort.
However, the operation of these AI-driven recommender systems involves substantial energy consumption, posing
a challenge to sustainability goals. The computational demands of data processing, model training, and deployment
contribute significantly to carbon emissions. Addressing this issue is critical in the context of the United Nations
Sustainable Development Goals (UNSDGs). This paper emphasizes the need for energy-efficient recommender systems
that minimize environmental impact while maintaining high performance.
This paper makes several contributions to the understanding of sustainable recommender systems in e-commerce.
First, we provide examples of recommender systems across different e-commerce applications. Second, we analyze how
each system enhances revenue. Third, we map these applications to a taxonomy of implementation methods. Fourth,
we examine the user effort required to obtain recommendations. Fifth, we offer suggestions for new, energy-efficient
recommender system applications.
The paper is valuable to two primary audiences: academics studying recommender systems in e-commerce and
practitioners implementing these systems. For academics, the examples and taxonomies provide a foundational frame-
work for future research. For practitioners, the paper offers guidance on selecting and implementing recommender
systems that align with both business goals and sustainability objectives.

2 Recommender System Examples

In the following section, we present six e-commerce businesses that utilize one or more variations of recommender
system technology. For each site and variation, we provide a brief description of the system’s features. In later
sections, we refer to these examples to explain the types of recommendations provided, the technologies used, and the
information gathered. These examples are alphabetized for organizational purposes and were accurate as of May 31,
1999. Due to the rapid evolution of the internet, some details may have changed.

2.1 Amazon.com
We focus on the recommender systems in the book section of Amazon.com. The ”Customers who Bought” feature on
each book’s information page offers two recommendation lists: one suggests books frequently purchased by customers
who bought the selected book, and the other recommends authors whose books are frequently purchased by customers
who bought works by the selected book’s author. This system personalizes the shopping experience and enhances
customer engagement by suggesting relevant products.

Eyes:
The Eyes feature allows customers to be notified via email about new items added to the Amazon.com catalog.
Customers can create requests based on author, title, subject, ISBN, or publication date information, using both
simple and complex Boolean-based criteria (AND/OR) for notification queries. These requests can be directly entered
from any search results screen, creating a persistent notification based on the search.

Amazon.com Delivers:
Amazon.com Delivers is a variation of the Eyes feature. Customers select checkboxes to choose from a list of spe-
cific categories or genres (e.g., Oprah books, biographies, cooking). Periodically, Amazon.com editors send email
announcements to notify subscribers of the latest recommendations in the subscribed categories.

Book Matcher:
The Book Matcher feature allows customers to provide direct feedback about books they have read. Customers
rate books on a 5-point scale from “hated it” to “loved it.” After rating a sample of books, customers can request
recommendations for books they might like. The system presents a selection of non-rated texts that correlate with
the user’s indicated preferences. Customers can further refine recommendations by rating these suggested books.

2
Customer Comments:
The Customer Comments feature provides text recommendations based on other customers’ opinions. Each book’s
information page includes a list of 1-5 star ratings and written comments from customers who have read and reviewed
the book. Customers can incorporate these recommendations into their purchase decisions.

2.2 CDNOW
Album Advisor:
The Album Advisor feature on CDNOW™ (www.cdnow.com) operates in two modes. In single album mode, it
recommends ten other albums related to the one the customer is viewing. In multiple artist mode, customers enter
up to three artists, and the system recommends ten albums related to those artists.

My CDNOW:
My CDNOW enables customers to set up a personalized music store based on their preferred albums and artists.
Customers indicate which albums they own and their favorite artists. Purchases from CDNOW are automatically
added to the ”own it” list. Customers can later distinguish between “own it and like it” and “own it but dislike it.”
The system predicts six albums the customer might like based on their collection and feedback. Customers can provide
further feedback to refine recommendations.

2.3 eBay
Feedback Profile:
The Feedback Profile feature on eBay.com™ (www.ebay.com) allows buyers and sellers to contribute feedback on their
transactions. Feedback includes a satisfaction rating (satisfied/neutral/dissatisfied) and a specific comment. Buyers
can view sellers’ profiles, which summarize ratings over different time periods (past 7 days, past month, past 6 months)
and provide an overall rating summary. Customers can browse individual ratings and comments for sellers.

2.4 Levis
Style Finder:
The Style Finder feature on the Levi Strauss™ (www.levis.com) website recommends clothing based on customer
preferences. Customers indicate their gender and rate items in three categories—Music, Looks, Fun—on a 7-point
scale from “leave it” to “love it” or choose “no opinion.” After rating a minimum number of items, customers receive
thumbnails of six recommended clothing items. Feedback can be provided to refine future recommendations.

2.5 Moviefinder.com
Match Maker:
Moviefinder.com’s Match Maker (www.moviefinder.com) allows customers to find movies with similar moods, themes,
genres, or casts to a given movie. From the information page of the selected movie, customers click the Match Maker
icon to receive a list of recommended movies and links to other films by the same director and key actors.

We Predict:
We Predict recommends movies based on customers’ previously indicated interests. Customers rate movies on a 5-
point scale (A to F). These ratings personalize the recommendation system, which provides tailored predictions for
unrated movies. Customers can also use Powerfind to search for top picks based on genre, directors, or actors and sort
results by personalized predictions or average customer ratings.

2.6 Reel.com
Movie Matches:
Similar to Amazon.com’s Customers who Bought feature, Reel.com’s Movie Matches (www.reel.com) provides recom-
mendations on each movie’s information page. Recommendations include “close matches” and “creative matches,”
each consisting of up to a dozen hyperlinks to similar movies, with annotated descriptions of how the new movie relates
to the original.

3
Movie Map:
The Movie Map feature on Reel.com recommends movies based on syntactic features. Customers enter queries based
on genre, movie type, viewing format, and price, and request results be constrained to “sleepers” or “best of this
genre.” Recommendations are editor’s picks fitting the specified criteria.

2.7 Summary
Table 1 summarizes the applications, interfaces, recommendation technologies, and methods for finding recommen-
dations across the example applications. The first column names each application under its respective e-commerce
site. The second column describes the interface used for delivering recommendations. The third column details the
recommendation technology and required inputs. The fourth column explains how users find recommendations using
the application. Each column is discussed in the paper, highlighting the role of each feature in supporting sustainable
and efficient recommender systems for e-commerce.

3 Recommendation Interfaces and Revenue Generation

An old proverb states that there is “more than one way to skin a cat.” Similarly, there are multiple ways to dis-
play recommendations to customers, and the chosen method often depends on the e-commerce site’s goals. Here,
we examine seven recommendation interfaces and how each helps the site generate revenue. While some methods
have roots in traditional commerce, each leverages the strengths of the electronic medium to provide more powerful
recommendations.

Browsing:
In traditional commerce, a customer might walk into a video store and ask the clerk to recommend “a comedy from the
50s.” Ideally, the clerk would suggest several movies, and the customer could browse the box covers to see which ones
appealed to them. However, the quality of these recommendations depended on the clerk’s knowledge of an enormous
range of movies. Reel.com enhances this process with its Movie Map feature by combining recommendations from
several clerks/editors, ensuring higher quality regardless of the query parameters. Additionally, recommendations
come with immediate links to the items, eliminating the need to search the store for obscure titles. This recommended
browsing converts browsers into buyers by helping users narrow down their choices and feel more confident in their
purchase decisions.

Similar Item:
The similar item recommendation is an adaptation of traditional commerce techniques. Systems like Reel.com’s Movie
Matcher, Amazon.com’s Customers who Bought, and one version of CDNOW’s Album Advisor expose customers
to items they may have forgotten or were unaware of. In e-commerce, these systems allow for more specific and
personalized recommendations based on items the customer has shown interest in, increasing exposure to the product
line and potentially boosting sales per order.

Email:
Email recommendations extend traditional direct mail techniques. Amazon.com’s Eyes feature notifies customers as
soon as an item becomes available, attracting them to the store before other stores can. Both Eyes and Amazon.com
Delivers keep customers aware of the site and items they might have missed. Customers appreciate these recommen-
dations because they help them stay informed about new items they are interested in purchasing. These features help
the site make money by increasing customer loyalty and the frequency of return visits.

4
 Figure 1: Table 1: Recommender System Examples

Text Comments:
Increasingly, sites provide recommendations based on text comments from other customers. Amazon.com’s Customer
Comments and eBay’s Feedback Profile streamline the collection of customer opinions, allowing potential buyers to
browse these comments. This helps convert browsers into buyers and increases site loyalty by providing impartial
information on products and sellers. Customers are more likely to return to a site they trust.

Average Rating:
A simpler approach to “the word on the street” is the average rating feature. Instead of browsing text-based opinions,
customers can see aggregated numerical rankings. Features like Customer Comments and Feedback Profile provide a
quick check on the quality of an item through average ratings. This facilitates converting browsers into buyers and
enhances customer loyalty.

Top-N:
Amazon.com’s Book Matcher, Levi’s Style Finder, and My CDNOW, among others, use top-N lists for recommen-
dations. After learning about a customer’s preferences, these sites provide a personalized list of top unrated items.
This is akin to gathering all potentially interesting items for a customer in one place, eliminating distractions. This
approach converts browsers into buyers by increasing exposure to relevant products and aiding decision-making on
items the customer was unsure about.

5
Ordered Search Results:
A less restrictive variation of the top-N list is Ordered Search Results recommendations. Instead of limiting predictions
to a predefined number, ordered search results allow customers to view items based on their predicted interest.
Moviefinder.com’s We Predict feature sorts query results by the likelihood that the customer will enjoy the item,
helping convert browsers into buyers.

Figure 2: Figure 1: Recommendation Taxonomy

4 A Taxonomy for Mapping Applications to Recommendation Tech-

niques
In this section, we delve into the Recommendation Technology column of Table 1. We will outline a taxonomy of
recommendation types, describe different user inputs (italicized in the table), and focus on user-centric analysis. Our
taxonomy emphasizes the degree of automation and persistence in recommendations (Figure 3).

Key Dimensions
Automation Axis: This ranges from completely automatic to completely manual recommendations. For the cus-
tomer, automatic recommendations are generated without any explicit effort, appearing seamlessly during site inter-
action. Manual recommendations require the customer to actively seek out recommendations. Note that recommen-
dations considered manual from the customer’s perspective might still be generated by the site’s algorithms.
Persistence Axis: This spans from ephemeral to persistent recommendations. Ephemeral recommendations are
based solely on a single session, not recognizing the customer in subsequent visits. Persistent recommendations, on
the other hand, remember the customer’s preferences across sessions, enhancing personalized suggestions.

6
 Figure 3: Recommendation Taxonomy

4.1 Non-Personalized Recommendations

Non-personalized recommender systems provide recommendations based on average customer feedback, independent of
the individual user. These systems are automatic and ephemeral, offering the same suggestions to all customers without
tracking individual preferences. Examples include Amazon.com’s average customer ratings and eBay’s feedback profile.
Both systems provide automatic and ephemeral recommendations, though eBay’s feedback profile also includes manual
text comments.

4.2 Attribute-Based Recommendations

Attribute-based recommender systems suggest products based on specific attributes or categories. These recommenda-
tions are often manual, requiring customers to specify their preferences, and can be either ephemeral or persistent. For
instance, Reel.com’s Movie Map recommends movies based on selected categories, making it manual and ephemeral.
Amazon.com Delivers, requiring category selection but persisting over time, exemplifies manual and persistent recom-
mendations.

4.3 Item-to-Item Correlation

These systems recommend products based on items the customer has shown interest in. They can be automatic, if
based on customer behavior, or manual, if input is required. Typically, they are ephemeral, not relying on past session
data. Examples include Reel.com’s Movie Matches and Amazon.com’s Customers who Bought, both automatic and
ephemeral. CDNOW’s Album Advisor, requiring input, is closer to manual.

4.4 People-to-People Correlation

People-to-people correlation systems, also known as collaborative filtering, recommend products based on similarities
between users. These systems are mostly automatic but require initial user input and are persistent, learning over
time. Examples include Amazon.com’s Book Matcher and My CDNOW, where user preferences are inferred from
actions and ratings, making them persistent and nearly automatic.

4.5 User Inputs

Each recommendation technology requires specific user inputs:
Purchase data: Used by systems like Amazon.com’s Customers who Bought. Likert: Scale-based inputs utilized by
eBay’s Feedback Profile. Text: Written comments in Amazon.com’s Customer Comments. Editor’s choice: Selected
by human editors, as seen in Reel.com’s Movie Matches.

7
5 Finding Recommendations
Different methods exist for customers to access recommendations, varying in effort required:

5.1 Organic Navigation

Requiring minimal effort, recommendations appear naturally as part of the site navigation, as seen in Album Advisor
and Feedback Profile.

5.2 Request Recommendation List

Customers can request recommendations based on their recorded preferences, as in Book Matcher and Style Finder.

5.3 Selection Options

Customers interact with the system, choosing from predefined criteria, as with Amazon.com Delivers and Moviefinder.com’s
We Predict.

5.4 Keyword/Freeform
The most interactive method, requiring customers to provide specific keywords for recommendations, exemplified by
Amazon.com’s Eyes.

6 E-Commerce Opportunities
While many recommender systems are in use, opportunities for expansion remain. Sites can explore implicit negative
ratings from returns or ignored detail views, enhancing algorithm performance. Recommender systems can also explain
product types based on past likes, integrate diverse data types for subtle personalization, and create marketing segments
using collaborative filtering.
Moreover, recommender systems can evolve into marketing tools, offering personalized pricing or targeted offers
based on customer behavior. Collaborative efforts among non-competing sites can improve recommendations through
shared user data, provided privacy is maintained.

7 Conclusion
Joe Pine’s book Mass Customization outlines five fundamental methods for achieving mass customization. Each of
the first four can be facilitated by recommender systems:

• Customize services around standardized products and services: Recommender systems provide a per-
sonalized service, allowing E-commerce sites to sell their largely standardized products more effectively.
• Create customizable products and services: Recommender systems themselves can be tailored to meet the
specific needs of an E-commerce site.
• Provide point of delivery customization: Recommender systems can directly customize the delivery point
for an E-commerce site.
• Provide quick response throughout the value chain: Future recommender systems are anticipated to
predict product demand, facilitating early communication along the supply chain.

Recommender systems are pivotal for automating mass customization on E-commerce platforms. Their importance
is expected to grow, as modern businesses increasingly focus on the long-term value of customers [?]. E-commerce
sites will strive to maximize customer value by offering tailored pricing and services that foster the most beneficial
relationships. While this often benefits both customers and sites, it also poses ethical challenges in balancing the
interests of both parties.
There are various techniques for implementing recommender systems, which can be selected independently of their
revenue-generation strategy. E-commerce sites can first determine their revenue-enhancement approach, then choose
the desired level of persistence and automation, and finally select a recommender system technique that fits this profile.
Contrary to the common belief that the ultimate recommender system is fully automatic and ephemeral, our study
reveals that many E-commerce sites favor persistent systems that require some manual input. Persistent systems help
build customer relationships, as customers tend to return to sites where they have invested effort, thus increasing the

8
site’s “stickiness.” In contrast, purely manual systems are portable, allowing customers to obtain recommendations
from any site offering similar manual features.
From the customer’s perspective, the optimal recommender system may be fully automatic and ephemeral, pro-
viding freedom to explore various E-commerce sites. However, for E-commerce sites, the ideal system is likely to
be persistent and partially automatic, requiring some customer input to enhance stickiness while providing valuable,
input-based recommendations. We predict that most recommender systems will be operated by E-commerce sites,
emphasizing persistence and partial automation. Some systems, aimed at supporting customers, will be ephemeral
and fully automatic to minimize effort, or ephemeral and fully manual to maximize control and portability [?].
Recommender systems add value for both E-commerce sites and their customers. We hope our taxonomy, which
details the ways recommender systems generate revenue, their implementation methods, and future directions, will
inspire the creativity needed to develop the recommender systems of tomorrow.

8 Acknowledgments
We gratefully acknowledge the support of the National Science Foundation under grants IIS 9613960, IIS 9734442,
and DGE 9554517. Additional support was provided by Net Perceptions Inc., a company co-founded by Konstan and
Riedl, which sells a recommender system for E-commerce.

9 REFERENCES
Christopher Avery, Paul Resnick, and Richard Zeckhauser 1999. The Market for Evaluations. American Economic
Review 89(3): pp 564-584.

Marko Balabanovic and Yoav Shoham 1997. Fab: Contentbased, collaborative recommendation. Communications
of the ACM, 40(3): pp. 66-72.

Chumki Basu, Haym Hirsh, and William Cohen 1998. Recommendation as classification: using social and content-
based information in recommendation. In Proceedings of the 1998 Workshop on Recommender Systems, pages 11-15.

John S. Breese, David Heckerman, and Carl Kadie 1998. Empirical analysis of predictive algorithms for collabo-
rative filtering. In Proceedings of the 14th Conference on Uncertainty in Artificial Intelligence (UAI98), pp 43-52.

Will Hill, Larry Stead, Mark Rosenstein, and George Furnas 1995. Recommending and evaluating choices in a
virtual community of use. In Proceedings of ACM CHI’95 Conference on Human Factors in Computing Systems,
pages 194-201.

Joseph A. Konstan, Bradley N. Miller, David Maltz, Jonathan L. Herlocker, Lee R. Gordon, and John Riedl 1997.
GroupLens: Applying collaborative filtering to Usenet news. Communications of the ACM, 40(3): pp 77- 87.

Don Peppers and Martha Rogers 1997. The One to One Future : Building Relationships One Customer at a Time .
Bantam Doubleday Dell Publishing.

B. Joseph Pine II 1993. Mass Customization. Harvard Business School Press. Boston, Massachusetts

B. Joseph Pine II, Don Peppers, and Martha Rogers 1995. Do you want to keep your customers forever? Har-
vard Business School Review, 1995(2): pp. 103-114.

Frederick F. Reichheld and W. Earl Sasser, Jr 1990. Zero Defections: Quality Comes to Services. Harvard Busi-
ness School Review, 1990(5): pp. 105-111.

Frederick F. Reichheld 1993. Loyalty-Based Management. Harvard Business School Review, 1993(2): pp. 64-73.

Paul Resnick, Neophytos Iacovou, Mitesh Suchak, Peter Bergstrom, and John Riedl 1994. Grouplens: An open
architecture for collaborative filtering of netnews. In Proceedings of ACM CSCW’94 Conference on ComputerSup-
ported Cooperative Work, pp 175-186.

Badrul M. Sarwar, Joseph A. Konstan, Al Borchers, Jon Herlocker, Brad Miller, and John Riedl 1998. Using fil-
tering agents to improve prediction quality in the grouplens research collaborative filtering system. In Proceedings of
1998 Conference on Computer Supported Collaborative Work.

9
Ben Shneiderman 1997. Direct Manipulation for Comprehensible, Predictable, and Controllable User Interfaces.
Proceedings of IUI97, 1997 International Conference on Intelligent User Interfaces, Orlando, FL, January 6-9, 1997,
33-39.

Upendra Shardanand and Patti Maes 1995. Social information filtering: Algorithms for automating ”word of mouth”.
In Proceedings of ACM CHI’95 Conference on Human Factors in Computing Systems, pages 210-217.

10