MD850: e-Service Operations

1
MD850 e-Service Operations

• Analyzing Web Site Usage Recommender Systems

2
Overview

• Background
• Recommender Systems
• Types of Recommender Systems
• Design Recommendations for Recommender Systems
• Issues with Recommender Systems
• Conclusion

3
Background
4
Background

• Customer decision making is an important activity
  within many e-services
• e-Retail Which item should I consider/buy?
• Online Newspaper Which articles are of
  interest to me?
• Entertainment sites Which movies are of
  interest to me?
• B2B When companies buy this widget, what other
  items do they commonly buy? I also might need
  those supplementary items.

5
Background

• Decision technologies tend to improve the
  decisions made by online customers
• Allows customers to compare and make trade-offs
  between various attributes of items under
  consideration
• Increases customers satisfaction with e-service
• Potentially makes customers more sure about the
  appropriateness of their purchase

Sources Jedetski, et al., How Web Site Decision
Technology Affects Consumers, IEEE Internet
Computing, April 2002
6
Background

• Before comparing alternative products (services,
  issues, scenarios, etc.)
• Customer must be able to identify alternatives
• Customer would like to make sure that the
  alternatives are relevant to their problem at hand

7
Background

• Customers typically do not have the time to
  search across the WWW to identify a decision
  making process for their problem at hand
• Too expensive
• Dont know what data should be used to make the
  decision
• Dont know where to get the data
• Dont know how to process the data

8
Background

• The WWW itself is a huge source of data for
  decision making
• Advantages
• Huge volume of information
• Lots of independent opinions contributed by
  biased and unbiased experts and novices
• Drawbacks
• Unstructured
• Large amount, but perhaps not a large amount
  collected anywhere about your specific problem

9
Background

• Customers provide a source of data within your
  e-service
• Web page requests
• Transactional data
• Items bought
• Events carried out within e-service
• Demographic data
• Opinions, experiences, beliefs

10
Background

• For the e-service designer, decision technologies
  provide a means for
• Enhancing the service product
• Differentiating your service experience from
  competitors
• Facilitating improved decision making of your
  customers
• Improving customer satisfaction

11
Background

• Decision technologies also provide several
  challenges for e-service design
• They often require collection and processing of
  data
• Must design processes for collecting data from
  system
• Must have statistical help available
• They become a programmed module within the
  service system
• Programmers must know how to implement the
  decision technology
• Poor implementation of the technology can hurt
  the speed and availability of your overall
  e-service

12
Background

• Types of e-service decision technology
• Identification of Alternatives
• Product hierarchy
• Search systems
• Recommender systems
• Suggests options to customer based on gathered
  information about decision space
• Comparison of Alternatives
• Recommender systems
• Data that the recommendation was based on can be
  presented
• Compensatory systems
• Allows decision maker to trade off attributes
  against one another
• Non-compensatory systems
• Eliminates bad options by setting minimum
  acceptable thresholds for certain (or all)
  attributes of options under consideration

Sources Jedetski, et al., How Web Site Decision
Technology Affects Consumers, IEEE Internet
Computing, April 2002
13
Recommender Systems
14
Recommender Systems

• Common Recommendation Methods
• Word of mouth
• Friends opinions and advice
• Recommendation letters
• Product reviews
• Movie reviews
• Book reviews
• Music reviews
• Entertainment reviews
• Surveys of an industry (Zagats Restaurant Guide)

15
Recommender Systems

• Common Recommendation Methods
• Based on low number of opinions
• Word of mouth
• Letters of recommendation
• Newspaper reviews (books, movies, etc.)
• Based on large number of opinions
• Zagats Restaurant Guides
• Regional paper (i.e., Boston Phoenix, Minneapolis
  City Pages) reader voting for Best Places To

16
Recommender Systems

• As the number of recommendations increases, the
  information provided by the recommendations
  should become better (more reliable, less biased)
• Letters of Recommendation
• If only one is required, applicant can easily
  choose person who is assured of giving a good
  recommendation
• If 3 are required, improves chance of observing
  negative information about applicant. Conversely,
  absence of negative letter provides stronger
  belief in positive information

17
Recommender Systems

• Economies of Scale
• The more data collected, the better the
  information about the recommendation
• As the sample size N approaches infinity, the
  shape of the distribution approaches the real
  distribution
• The larger the number of recommenders, the more
  likely it is that youll find recommenders
  similar to you
• If data are only collected for a population that
  does not represent you, it does not help you out

18
Recommender Systems

• Shortcomings of common (person-to-person)
  recommendation systems
• Small number of recommendations collected
• Recommendations are more likely to be biased
• Small number of recommenders
• Less likely that you have the same
  characteristics as the recommender

19
Recommender Systems

• Recommender Systems
• Recommender systems assist and augment the
  natural social process of making decisions
• Built using information technology
• People provide inputs to the system
• Recommendation/ratings of products, etc.
• Content available from various sources
• System aggregates inputs
• System makes suggestions to persons based on
  information collected from other people/sources

20
Recommender Systems

• Friends Recommendations vs. Recommender Systems
• People tend to rate recommendations from their
  friends as better than recommendations made by
  online recommender systems
• Human recommenders tend to recommend items that
  remind the person about something they are
  already aware of
• Online recommenders tend to provide new and
  unexpected items about which they were not
  aware
• Users tend to like the breadth of recommendations
  made available by online systems friends cant
  know as many items

Sources http//www.sims.berkely.edu/sinha/Recomm
enders.html
21
Recommender Systems

• Identifying Successful Recommendations
• Success can be defined in different ways
• E-commerce success for a specific customer, the
  system identifies a product/service that the
  customer is likely to buy
• Exploration/Learning success the system helps
  users to explore their tastes

Sources http//www.sims.berkely.edu/sinha/Recomm
enders.html
22
Recommender Systems

• Desirable Performance Characteristics
• Scalable over very large customer bases
• Scalable over very large product catalogs
• Sub-second processing time to generate a set of
  recommendations
• Able to react immediately to changes in user data
• Makes compelling recommendations for all users
  regardless of what theyve purchased, rated, or
  done previously

Sources Linden, G., et al., Amazon.com
Recommendations Item-to-Item Collaborative
Filtering, IEEE Internet Computing, Jan/Feb, 2003
23
Recommender Systems

• Examples of Recommender Systems
• Music
• Amazon.com
• MediaUnbound.com
• MoodLogic.com
• CDNow.com
• SongExplorer.com

24
Recommender Systems

• Examples of Recommender Systems
• Movies
• Amazon.com
• Moviecritic.com
• Reel.com
• Sepia Video Guide
• MovieFinder.com
• Morse

25
Recommender Systems

• Examples of Recommender Systems
• Books
• Amazon.com
• RatingZone.com QuickPicks
• Sleeper (pmetrics.com)

26
Recommender Systems

• Inputs to the System
• Recommenders differ widely in the number of
  inputs they require before they will start making
  recommendations (generally, between 1 and 30)
• Ratings inputted
• Open-ended phrases, words, etc.
• Ratings on a Likert Scale (5 or 7 point scale)
  Like a lot, Like, Unsure, Dislike,
  Dislike a lot
• Binary liking Like (1)/Do Not Like (0)
• Hybrid rating process combination of the above

27
Recommender Systems

• Inputs to the System
• Web Usage Mining
• Customer activities within the web site can be
  mined from web site transaction logs
• Data collection
• Data preparation
• Discovery of usage profiles
• Using usage profiles as a basis for making
  recommendations

28
Recommender Systems

• Processing of Inputs
• Database of records for individuals
• Fields are
• items they have bought, or that they own
• ratings of items (books, movies, Web documents,
  etc.) they have rated
• Information in database may be pre-processed
  offline to create tables of similarity measures
  between customers, products, etc.

29
Recommender Systems

• Generating Recommendations
• Vector of a customers ratings is compared to the
  vectors provided by other users
• People with similar opinions can be discovered
• Recommendations for the customer is based on
  observed patterns for the prior customers

30
Recommender Systems

• General Types of Recommendations
• Prediction
• a prediction that a user will like a specific
  item
• Top-N Recommendation
• a list of N items that a users is most likely to
  choose or be interested in
• Top-M Users
• predict a list of M users who will like a
  specific item the most

31
Types of Recommender Systems
32
Types of Recommender Systems

• Manual decision rule systems
• allow web site administrators to specify rules
  based on user demographics or static profiles of
  users
• Ex Broadvision

Sources Wei, C., et al., A Survey of
Recommendation Systems in Electronic Commerce,
e-Service, 2002
Sources Linden, G., et al., Amazon.com
Recommendations Item-to-Item Collaborative
Filtering, IEEE Internet Computing, Jan/Feb, 2003
Sources Mobasher, B., et al., Automatic
Personalization Based on Web Usage Mining,
Communications of the ACM, 2000
33
Types of Recommender Systems

• Popularity-based recommender systems
• identifies the most popular items within a given
  community
• provides a means for finding out what one should
  be paying attention to among ones peers

Sources Wei, C., et al., A Survey of
Recommendation Systems in Electronic Commerce,
e-Service, 2002
Sources Linden, G., et al., Amazon.com
Recommendations Item-to-Item Collaborative
Filtering, IEEE Internet Computing, Jan/Feb, 2003
Sources Mobasher, B., et al., Automatic
Personalization Based on Web Usage Mining,
Communications of the ACM, 2000
34
Types of Recommender Systems

• Collaborative filtering
• take explicit information in the form of user
  ratings or preferences, and through a correlation
  engine, return information that is predicted to
  closely match the users preferences
• Ex Firefly, NetPerceptions

Sources Wei, C., et al., A Survey of
Recommendation Systems in Electronic Commerce,
e-Service, 2002
Sources Linden, G., et al., Amazon.com
Recommendations Item-to-Item Collaborative
Filtering, IEEE Internet Computing, Jan/Feb, 2003
Sources Mobasher, B., et al., Automatic
Personalization Based on Web Usage Mining,
Communications of the ACM, 2000
35
Types of Recommender Systems

• Collaborative Filtering
• A major type of recommender system
• Social intelligence
• Many of us behave similarly, due to having
  similar preferences
• Human minds have ability to infer tastes from
  wardrobe and prior tastes
• Hes a punk rocker. Hes weird. He must like The
  Ramones, The Replacements, Husker Du, and
  Nirvana.
• If we can identify behaviors, we can help
  individuals find things of use to them

36
Types of Recommender Systems

• Collaborative filtering helps with decisions that
  are related to human tastes
• Books -- preferences for certain authors
• Music -- tastes in music
• Movies -- tastes in entertainment
• News -- tastes in stories
• Restaurants -- tastes in food
• Website Pages -- tastes in information content
• Cross-Selling -- common related purchases
  (complements)

37
Types of Recommender Systems

• Content based filtering
• calculate the similarity between document (or
  product) content and information (explicit or
  implicit) in personal profiles of users
  recommend documents having high similarity
• Ex WebWatcher

Sources Wei, C., et al., A Survey of
Recommendation Systems in Electronic Commerce,
e-Service, 2002
Sources Linden, G., et al., Amazon.com
Recommendations Item-to-Item Collaborative
Filtering, IEEE Internet Computing, Jan/Feb, 2003
Sources Mobasher, B., et al., Automatic
Personalization Based on Web Usage Mining,
Communications of the ACM, 2000
38
Types of Recommender Systems

• Cluster modeling systems
• uses cluster analysis (offline) to segment
  customers into related groups
• each customer is allocated to the segment that is
  most similar to them
• online recommendations are generated based on
  what products the individuals in a segment have
  purchased

Sources Wei, C., et al., A Survey of
Recommendation Systems in Electronic Commerce,
e-Service, 2002
Sources Linden, G., et al., Amazon.com
Recommendations Item-to-Item Collaborative
Filtering, IEEE Internet Computing, Jan/Feb, 2003
Sources Mobasher, B., et al., Automatic
Personalization Based on Web Usage Mining,
Communications of the ACM, 2000
39
Types of Recommender Systems

• Search-based methods
• treat the recommendation as a search for related
  items constructs a search query to find other
  items by the same author, manufacturer, musician,
  keyword, etc.

Sources Wei, C., et al., A Survey of
Recommendation Systems in Electronic Commerce,
e-Service, 2002
Sources Linden, G., et al., Amazon.com
Recommendations Item-to-Item Collaborative
Filtering, IEEE Internet Computing, Jan/Feb, 2003
Sources Mobasher, B., et al., Automatic
Personalization Based on Web Usage Mining,
Communications of the ACM, 2000
40
Types of Recommender Systems

• Association-based Recommendation Approach
• relies on user preferences to identify items
  frequently found in association with items that a
  user has chosen
• use item attributes to identify other items that
  are similar
• if a customer has chosen an item, suggest that
  they may like the similar items

Sources Wei, C., et al., A Survey of
Recommendation Systems in Electronic Commerce,
e-Service, 2002
Sources Linden, G., et al., Amazon.com
Recommendations Item-to-Item Collaborative
Filtering, IEEE Internet Computing, Jan/Feb, 2003
Sources Mobasher, B., et al., Automatic
Personalization Based on Web Usage Mining,
Communications of the ACM, 2000
41
Types of Recommender Systems

• Demographics-based Approach
• recommends items to a user based on the
  preferences of other users with similar
  demographics

Sources Wei, C., et al., A Survey of
Recommendation Systems in Electronic Commerce,
e-Service, 2002
Sources Linden, G., et al., Amazon.com
Recommendations Item-to-Item Collaborative
Filtering, IEEE Internet Computing, Jan/Feb, 2003
Sources Mobasher, B., et al., Automatic
Personalization Based on Web Usage Mining,
Communications of the ACM, 2000
42
Types of Recommender Systems

• Reputation-based Approach
• identifies users that a customer respects (i.e.,
  opinion leaders, reviewers, experts)
• uses the opinions of these selected users as a
  basis for recommendations to the customer

Sources Wei, C., et al., A Survey of
Recommendation Systems in Electronic Commerce,
e-Service, 2002
Sources Linden, G., et al., Amazon.com
Recommendations Item-to-Item Collaborative
Filtering, IEEE Internet Computing, Jan/Feb, 2003
Sources Mobasher, B., et al., Automatic
Personalization Based on Web Usage Mining,
Communications of the ACM, 2000
43
Design Recommendations for Recommender Systems
44
Design Recommendations for Recommender Systems

• Number of items users must rate to receive
  recommendations
• Users tend not to mind providing more input to
  the recommender system in order to get better
  recommendations
• Trust in system
• Previously liked recommendations play an
  important role in establishing trust in the
  system
• If a system is recommending a lot of items a
  person has liked previously, it provides trust in
  being able to like the newly recommended items
• System transparency
• Users like to understand the logic behind the
  recommendations being made
• Users want to know Why was an item recommended?

Sources http//www.sims.berkely.edu/sinha/Recomm
enders.html
45
Design Recommendations for Recommender Systems

• Level of detail about recommended item
• A recommendation needs to be backed up by more
  information about the recommended item
• Users find it difficult to judge a recommended
  item if the system provides few details
• Ability to filter recommendations by genre
• After recommendations are made, users like to be
  able to choose to see only the recommendations
  from a certain genre (e.g., hard rock) or product
  classification
• Interface matters when it gets in the way
• If system layout and navigation hurt the site,
  they will hurt the value of the recommender system

Sources http//www.sims.berkely.edu/sinha/Recomm
enders.html
46
Issues With Recommender Systems
47
Issues With Recommender Systems

• Capacity Management/Delivery Speed
• Recommender systems are often computationally
  expensive
• Expensive computations tend to slow down speed of
  e-service
• Common strategies for deciding how to collect and
  analyze recommendations, and how to integrate
  them into the recommendation making process
• Offline processing of data
• Periodic updating of data matrix used to make
  recommendations

48
Issues With Recommender Systems

• Free Rider problem
• Customer has an incentive to register and use
  recommender system (a public resource within
  the e-service)
• Saves them time
• Improves their decisions
• Customer may not have an incentive to contribute
  their own recommendations to the system
• Consumes time to add recommendations to system
• Information for other customers will not be
  improved if customer doesnt contribute their
  opinions/recommendations/ratings

Sources Resnick and Varian, Recommender
Systems, Communications of the ACM, March 1997
49
Issues With Recommender Systems

• Trying to Rig the System
• People/companies who are being rated have an
  incentive to make sure that the ratings are
  favorable toward them
• Individuals/companies may try to generate
• mountains of positive information about their own
  products/services
• mountains of negative information about their
  competitors products/services
• If this information is incorporated into the
  system, it leads to biased recommendations
• Customers, over time, will perceive the bias and
  stop using the recommender system

Sources Resnick and Varian, Recommender
Systems, Communications of the ACM, March 1997
50
Issues With Recommender Systems

• Personal Privacy
• People may not want their habits/views to be
  known
• Anonymity can improve the number of
  recommendations collected
• Some people like to get credit for their
  contributions to the system
• Participation under a pseudonym
• Attributed credit
• Ex Amazon and other sites have Top Reviewer
  categories

Sources Resnick and Varian, Recommender
Systems, Communications of the ACM, March 1997
51
Summary
52
Summary

• Decision technologies help customers make
  decisions in an e-service, and can thus improve
  customer satisfaction
• Recommender systems use customer information to
  improve decision making of other customers
• Several different types of recommender systems
• Must be careful in choosing recommender, to make
  sure that recommendations are appropriate
• Must consider speed and scalability of
  recommender algorithm