Recommendation Systems

1
Recommendation Systems

• Jeremy Bonjio
• Eric McGregor

2
Overview

• Recommender Systems
• The Netflix Prize
• Current Netflix Prize Leaders
• Collaborative Filtering
• The Bellkor CF Algorithm
• Our Semi-Clone Implementation
• References

3
Recommender Systems

• Definition
• Recommender systems (RS) analyze patterns of
  user interest in items to provide personalized
  recommendations of items that will suit a users
  taste.1

4
Benefits of RS

• Value for Businesses
• may decrease use of system resources
• may improve inventory management
• Value for Customers
• may decrease time required to make a selection

5
Common Types of RS

• Content Based Approach
• create profiles of each product and user
• requires gathering of external information
• Collaborative Filtering (CF)
• relies only on past user behavior
• e.g. previous transactions or user ratings
• does not require profiles on items or users
• analyzes relationships between users and
  similarities among products

6
Online RS

• Amazon.coms Item Recommender
• Yahoo! Musics Online Music Recommender
• Netflixs Cinematch Movie Recommender

7
The Netflix Prize Contest Goal

• The Netflix Prize seeks to substantially improve
  the accuracy of predictions about how much
  someone is going to love a movie based on their
  (and others past) movie preferences.2

8
Netflix Prize Contest

• Input
• a set of user ratings for a set of movies
• a.k.a. the training data
• a set of (user-id, movie-id) pairs which are not
  given in the training data
• a.k.a. the qualifying test
• Goal
• to determine for each pair (u, m) in the
  qualifying test, the rating that u actually gave
  for m.

9
Judging the Netflix Prize Results

• RMSE (root mean squared error) is computed as the
  square root of the averaged squared difference
  between each prediction and the actual rating
• Predictions must be 10 greater than Cinematch
  using RMSE in order to win Grand Prize.

10
Training Data Set

• Consists of more than 100 million ratings (1-5)
• from over 480 thousand randomly-chosen, anonymous
  customers
• on nearly 18 thousand movie titles
• The date of each rating is provided
• The title of each movie is provided
• The year of release for each movie is provided
• No other customer or movie information is provided

11
Qualifying Test Set

• Contains over 2.8 million customer/movie id pairs
  
• with rating dates
• but with the ratings withheld.
• These pairs were selected from the most recent
  ratings from a subset of the same customers in
  the training data set, over a subset of the same
  movies.

12
Current Netflix Prize Leader

• Team Bellkor
• consists of Bob Bell and Yehuda Koren or ATT
  Research
• in the Statistics and the Information
  Visualization departments. 3
• Collaborative Filtering Approach 1

13
Collaborative Filtering

• Can be viewed abstractly as
• Missing Value Estimation Problem
• note more missing values - the harder the
  problem

14
Forms of Collaborative Filtering

• neighborhood approach
• a.k.a. kth Nearest Neighbor or kNN
• most common
• used by Bellkor
• factorization approach

15
kNN Methods of CF

• To predict rating of item i by user u
• Identify set of items (neighbors) that tend to be
  rated similarly to i
• each neighbor must have been rated by u
• Assign each neighbor a weight
• Predict rating rui ?j?N(i,u) wij ruj

16
Major Components of kNN Method

• Data Normalization
• simple solution adjust for varying mean ratings
  across users (and/or items)
• more complex removing effects
• Neighborhood Selection
• Pearson Correlation Coefficients
• cosine similarity
• Determination of Interpolation Weights for Each
  of k-Neighbors
• simple solution use Pearson Correlation
  Coefficient
• more complex Bellkor method
• Calculate predictions
• for user u and item i, rui ?j?N(i,u) wij ruj

17
Bellkor Data Normalization

• Complex Normalization
• removal systematic effects
• e.g. systematic tendencies for some users to give
  higher ratings than others
• removal of characteristic effects
• e.g. ratings for some movies may fall over time
  but not be evident in the set of ratings
• removal of variable effects
• e.g. if the number of ratings for a movie is low
  the average rating may not necessarily be a good
  measurement

18
Bellkor Determination of Interpolation Weights

• More advanced then using Pearson Correlation
  Coefficients
• Accounts for interactions among neighbors
• Avoids triple counting
• e.g. Neighbor set containing trilogy
• Avoids over-fitting
• e.g. The user has not watched any movies similar
  to the item

19
Our Semi-Clone Implementation

• Normalization
• simply adjust for varying mean ratings across
  users
• Neighborhood Selection
• Pearson Coefficient Constants
• Determination of Interpolant Weights
• clone Bellkor process

20
References

• R.Bell and Y. Koren, ATT Research, Scalable
  Collaborative Filtering with Jointly Derived
  Neighborhood Interpolation Weights
• http//www.netflixprize.com
• http//www.research.att.com/volinsky/netflix