CSCI 7000 Modern Information Retrieval Jim Martin

1
CSCI 7000Modern Information RetrievalJim Martin

• Lecture 23
• 12/1/2008

2
Today 12/1

• Collaborative Filtering
• Recommender systems

3
Collaborative Filtering

• Filtering in the IR context generally means
  separating things (documents) into categories.
  That is, classification.
• Binary special case
• Yes/No. I care about it or I dont.
• As weve seen its typically done by training
  supervised or unsupervised systems based on the
  contents of the items being classified.
• By contrast, in collaborative filtering its done
  by looking primarily at the behavior of other
  users.

4
Recommender Systems

• Recommender systems are a specific application of
  collaborative filtering.
• Recommend a product (movie, book, music, etc.) to
  a user based on...
• That users past behavior with respect to other
  products
• Other users behavior with respect to other
  products
• Behavior here could mean purchases, or could
  mean explicit ratings

5
Example CF Technology iTunes Genius
6
Amazon

• Probably the most well-developed system.
• Multiple presentations to the user
• Recommendations for you
• People who bought X also bought Y
• Of the people who looked at X, some bought the
  following...
• Good recommendations drive sales

7
Netflix Prize

• Improve the Netflix recommender system by 10 and
  win 1M.

8
Basic Approach

• You have
• A bunch of items
• A bunch of users
• And data recording users behavior with respect to
  the items
• Ratings (ordinals or reals)
• Purchases (binary)
• So lets do what always do and make matrix. In
  this case
• A user x item matrix

9
Basic Approach

• As usual the matrix is mostly empty.
• These empty cells represent opportunities to make
  recommendations.

Items
Users
10
Basic Approaches

• Given an User x Item matrix you can take three
  basic approaches.
• User-based
• Similarity among users
• Item-based
• Similarity among items
• Model-based
• Fill in the missing cells with plausible values
  based on the structure of the existing matrix

11
User Based Methods

• For a given user of interest
• Ie. A customer logged in to Amazon
• Compute the similarity of all users to the
  current user and select a subset to use as
  predictors
• Normalize ratings across users
• Make predictions for current user by using a
  weighted combination of other user ratings.

12
User-Based Methods

• Pearson correlation
• Ranges from -1 to 1
• 1 means the ratings between users are perfectly
  correlated
• -1 means theyre perfectly anti-correlated
• 0 means that theres no observable correlation.

13
Item-Based Methods

• For the user of interest, select a subset of
  highly rated items from their vector.
• Compute the similarity score between those
  vectors and other currently empty entries for
  that user.
• Recommend the N closest unrated entries.
• Basically the same method. Just flips users and
  items.

14
Model-Based Methods

• Treat the missing entries like 0s in a language
  model, or a probabilistic model.
• One approach is to just use LSA to do a
  dimensionality reduction on the matrix.

15
Evaluation

• Given a training set, the typical scenario is to
  use a held-out scheme
• Typically based on holding out a rating, not an
  item or a user
• For example
• Hold out 1 rating per user (row) and predict the
  missing entries.
• Compare predicted values to known values.
• Average over the users
• Or
• Sample K ratings per user and predict the missing
  values.

16
Datasets

• www.grouplens.org

17
Lots of Issues

• Fundamental problem underlying all of these
  approaches and the whole paradigm...
• Stationarity assumption
• Basically that things arent changing in ways
  that matter with respect to the probabilities in
  question
• But of course thats silly tastes change and
  people change all the time.

18
Lots of Issues

• The current way of casting the problem ignores
  the temporal dimension.
• A users future behavior is based their past
  behavior in funny ways that the current model
  doesnt handle well.
• For example, it doesnt capture marginal utility.
  
• How many intro books on a given programming
  language does one person need?
• The fact that a system might accurately predict
  my hypothetical rating of an item is not a
  prediction about whether or not I care about that
  item at all.

19
Lots of Issues

• In many settings, various kinds of metadata is
  available
• Demographic information about users (gender, age,
  location, etc.)
• Information about the products (price, category,
  prices of similar items, age, features, etc.)
• Incorporating such information into the basic
  scheme is problematic.

20
Lots of Issues

• Privacy
• Storing data on user choices is a tricky business
  with obvious privacy concerns
• Most efforts rely on deidentification
  techniques...
• Replace any real world identifier with a unique
  meaningless term...
• But that doesnt always work
• AOL fiasco
• Netflix partially hacked

21
Readings for Collaborative Filtering

• Read Chapter 2 of Collective Intelligence
• And the Adomavicius et al and the McNee et all
  papers linked to on the web page.

22
Quiz

• Quiz is Wednesday
• Focus on material since last quiz
• Information extraction
• Sentiment and opinion
• Recommender systems
• But you should remember the earlier stuff
• Retrieval, classification and clustering