What is a Recommender System

1
What is a Recommender System ?
September, 2006 Juntae Kim Dept. of Computer
Engineering, Dongguk University Seoul, Korea
2
Contents

• Concepts
• Types of recommendations
• Brief history
• Issues and future directions

3
Concepts

• Definition Wikipedia
• Recommendation systems are programs which attempt
  to predict items (movies, music, books, news, web
  pages) that a user may be interested in, given
  some information about the user's profile.

4
Concepts
5
Concepts
6
Contents

• Concepts
• Types of recommendations
• Brief history
• Issues and future directions

7
Non-Personalized

• Use popularity
• Peoples ratings, sales data
• Accumulate rating data (or transaction data)
• Recommend items with highest average rating
• Exgt Provide a list of top-10 most popular albums
• Exgt User specifies that he/she likes Jazz. The
  system provides a list of top-10 most popular
  Jazz albums

8
Demographic

• Use user features
• Age, gender, area code, preferred category,
• Find users with similar features
• K-nearest neighbor with feature vectors
• Recommend items that are preferred by similar
  users
• Weighted average of ratings

9
Demographic

• Computing similarity
• Define distance functions
• Computing prediction
• Weighted average of neighbor users ratings
• rating of user u to item j
• predicted rating of user a to item j

10
Demographic

• Advantage
• New user can get recommendations without
  providing preference information
• Problems
• Insufficient information
• Highly personalized recommendation can not be
  achieved

11
Content-Based

• Use item features
• Keywords, author, genre,
• Find items with similar features compared to the
  known preferred items (or to the user profile)
• Cosine similarity between keyword vectors
• Recommend items with highest similarity

12
Content-Based

• Computing similarity
• Cosine similarity between keyword vectors Vi, Vj
• Computing prediction
• Average of predicted ratings
• Set of items preferred by the user a

13
Content-Based

• Advantage
• It performs well in terms of relevancy (to users
  interest) if items can be properly represented as
  a set of features
• Exgt Text ? set of keywords Good NLP techniques
  and IR theories
• Problems
• Content analysis is necessary for feature
  extraction
• Difficult to apply Multimedia contents (image,
  video, music)
• Quality of item can not be evaluated
• How relevant is the item? vs. how good is the
  item?

14
Collaborative User-Based

• Use user-item preferences
• Explicit user ratings, transaction data
• Find highly correlated users
• Pearson correlation between user rating data
• Recommend items that are preferred by those users
• Weighted average of ratings

15
Collaborative User-Based

• Computing similarity
• Cosine similarity
• Pearson correlation
• rating of user u to item i
• average rating of user u
• Set of items rated by both users a and u

16
Collaborative User-Based

• Computing prediction
• Weighted average of neighbor users ratings
• Set of users similar to user a

17
Collaborative User-Based

• Advantages
• Content analysis is not necessary
• Quality (not just relevancy) can be evaluated by
  collaboration
• Problems
• Cold-start at initial stage, users can not get
  good recommendations
• Scalability
• Sparsity
• Users with different opinion or unusual taste may
  not get good recommendations

18
Collaborative Item-Based

• Use user-item preferences
• Explicit user ratings, transaction data
• Find items that are highly correlated with the
  known preferred items
• Pearson correlation, vector similarity
• Recommend items with highest correlation

19
Collaborative Item-Based

• Computing similarity
• Cosine similarity
• Adjusted cosine similarity
• rating of user a to item i
• average rating of user a
• Set of users who rated both items i, j

20
Collaborative Item-Based

• Computing prediction
• Average of predicted ratings
• Set of items rated by user a

21
Collaborative An Example
Recommend items preferred by highly correlated
user U3 ? I5
22
Collaborative An Example
Recommend items highly correlated to preferred
items ? I5
23
Collaborative Item-Based

• Advantages
• No content analysis, evaluation of quality are
  same as user-based collaborative filtering
• Scalable
• Alleviate sparsity problem
• Problems
• Cold-start problem, peculiar user problem are
  same as user-based collaborative filtering

24
Model-Based

• Use user-item preferences, or user/item features
• Build a model for prediction
• Probabilistic, decision tree, neural net, ...
• Recommend items by using the model
• Computing probability, applying decision rules,

25
Model-Based

• Probabilistic model
• Probability of a rating value given known ratings
• Set of items rated by user a
• The probability can be estimated by Bayesian
  Network learning
• Computing prediction
• Expected rating

26
Model-Based

• Advantages
• Based on statistical theory
• Relatively high performance has been reported
• Accuracy
• Time
• Problems
• It needs large amount of data to build an
  accurate model
• Not incremental. The model must be re-build when
  the data is updated

27
Contents

• Concepts
• Types of Recommendations
• Brief history
• Issues and future directions

28
Early Works

• Tapestry Goldberg et al., 1992
• Filtering emails by Collaborative Filtering
• Accepts annotations of users for electronic
  documents
• Each user constructs a filter by using a query
  language to search the annotations
• Ringo Shardanand and Mae, 1995
• Music recommendation by Social Information
  Filtering
• Collects users profiles (ratings 1-7 to
  artists), and perform collaborative filtering
• Started July, 1994. 2100 users, 3000 artists,
  9000 albums
• The recommendation result shows better MAE
  compared to recommendation with average rating
• ? FireFly Networks, Inc.

29
Early Works

• GroupLens Resnick et al., 1994 Konstan et al.,
  1997
• Recommends Usenet news (later, movies)
• Collects users ratings (1-5 ratings to
  articles), and performs collaborative filtering
• 7 weeks trial in Feb. 1996. 250 users, 22000
  articles, 47000 ratings
• The recommendation result shows higher
  correlation between prediction and actual rating
  than recommendation based on average rating
• ? NetPerceptions, Inc.

30
Recommending Web pages

• Content-based
• Syskill Webert Pazzani and Billsus, 1997
  Given users preferred pages, builds a user
  keywords profile. Recommendations are made by
  similarity computation between the user profile
  and Web pages
• Content Collaborative
• Fab Balabanovic and Shoham, 1997 Similar users
  are identified based on user keyword profile.
  Recommendations are made based on similar users
  ratings
• Siteseer Rucker and Polanco, 1997 Similar
  users are identified based on bookmark listings.
  Recommendations are made based on similar users
  bookmark items
• PHOAKS Terveen et al., 1997 Recommend valuable
  web resources by analyzing newsgroup messages
  containing URL recommendation

31
Collaborative Filtering

• Various algorithms and modifications
• Breese et al., 1998 Herlocker et al., 1999
  Presents various modifications to collaborative
  filtering for performance improvement
• Sarwar et al., 2000 Applies dimensionality
  reduction (SVD) to overcome sparsity problem
• Goldberg et al., 2001 Jester - collaborative
  joke recommendation. Maps users into 2D space by
  using principal component analysis
• Integrating contents
• Basu et al., 1998 Uses collaborative and
  content-based features together in a rule-based
  classifier for movie recommendation
• Pazzani, 1999 Performs collaboration via
  content profile
• Good et al., 1999 Specialized content analysis
  agent participates in collaborative filtering
• Melville et al., 2002 Augments users rating
  vector with additional content-based predicted
  ratings

32
Collaborative Filtering

• Item-based collaborative filtering
• Sarwar et al., 2001 Item-based collaborative
  filtering algorithm. Computes correlation between
  items
• Linden el al., 2003 Describes Item-to-item
  collaborative filtering algorithm used in
  Amazon.com
• Deshpande and Karypis, 2004 Presents
  item-based top-N recommendation algorithm and
  experiment with 8 different dataset (purchase
  transactions, MovieLens, etc.)
• Amazon.com
• Recommends books, CDs, DVDs, etc.
• Item-based collaborative filtering

33
Probabilistic Models

• Probabilistic model-based recommendation
• Ungar and Foster, 1998 Builds clusters of
  users and movies with K-means and Gibbs sampling
• Chien and George, 1999 Presents a Bayesian
  model
• Shani et al., 2002 Introduces Markov decision
  process model
• Unified probabilistic frameworks
• Condliff et al., 1999 Bayesian mixed effect
  model that incoporate user ratings, user
  features, and item features
• Popescul et al., 2001 Schein et al., 2002
  Hofmann, 2003 Unified probabilistic framework
  for combining collaborative and content-based
  recommendation based on probabilistic latent
  semantic analysis
• ChoiceStream, Inc.
• Recommends online contents and products
• Bayesian model based on contents and user
  behavior

34
Academic Workshops

• 1996 Berkeley Collaborative Filtering Workshop
• 1997 Communications of ACM Recommender Systems
• 1998 AAAI Workshop on Recommender Systems
• 1999 SIGIR Recommender Systems Workshop
• 2001 SIGIR Workshop on Recommender Systems
• 2006 ECAI Workshop on Recommender Systems

35
Applications SamsungMall
36
Applications Amazon.com
37
Applications CDNow
38
Applications MusicStrands
39
Applications AOL television
40
Contents

• Concepts
• Types of Recommendations
• Brief history
• Issues and future directions

41
Issues Future Directions

• Sparsity
• Very sparse user-item matrix degrades performance
• ? Transforming data augmented rating,
  dimensionality reduction
• Scalability
• Real-time computation with millions of
  users/items is needed
• ? Scalable algorithms Item-based CF,
  model-based system
• Cold-start
• Little rating data in initial stage. New user,
  new item problems
• ? Integrating demographic and content
  information
• ? Active learning

42
Issues Future Directions

• Implicit ratings
• It is hard to get large amount of explicit
  ratings
• ? Acquiring and analyzing implicit data
  purchase, listen, Web usage patterns like
  click sequence, time of reading, etc.
• ? Recommendation with positive-only data
• Dealing with multiple criteria
• Some applications have multiple rating criteria
• Exgt Hotel recommendation room, facility,
  service, location, food, etc.
• ? Incorporating optimization techniques
• Context-dependant recommendation
• Recommendation may vary according to time, place,
  company, etc.
• Exgt Recommendation of music at a party vs. in the
  bed room
• ? Introducing context tags. Probabilistic model
  with contextual variables

43
Issues Future Directions

• Malicious usage
• Some users lie to the recommender systems in
  order to have their items recommended (shilling)
• Exgt Hire users to give good rating to their
  product
• Exgt Run software agents to produce high traffic
  to specific items
• ? Detection of shilling. Protecting system from
  shilling attack
• Privacy
• Users personal taste or private actions
  (collected for implicit rating) can be revealed
• Open to system administrator
• Revealed by analyzing recommendation results
• ? Providing privacy protection mechanism

44
Thank you !