Item-Item Collaborative Filtering

In this assignment, a simple implementation of item-item collaborative filtering is done.

Implementing Item-Item Collaborative Filtering

The classes used are:

SimpleItemItemModelBuilder: Builds the item-item model from the rating data

SimpleItemItemScorer: Scores items with item-item collaborative filtering

SimpleItemBasedItemScorer: Finds similar items

The provided SimpleItemItemModel class stores the precomputed similarity matrix.

Computing Similarities

The SimpleItemItemModelBuilder class computes the similarities between items and stores them in the model. It also needs to create a vector mapping each item ID to its mean rating, for use by the item scorer. The following configuration decisions are used:

• Each item rating vector is normalized by subtracting the item's mean rating from each rating prior to computing similarities.
• Cosine similarity are used between normalized item rating vectors.
• Only neighbors with positive similarities (> 0) are stored.

The similarity matrix is in the form of a Map from Longs (items) to Long2DoubleMaps (their neighborhoods). Each Long2DoubleMap stores a neighborhood, where each neighbor's id (the key) is associated with a similarity score (the value).

Scoring Items

The SimpleItemItemScorer class uses the model of neighborhoods to actually compute scores. The items are scored using the weighted average of the users' ratings for similar items.

At most 20 neighbors are used to score each item; if the user has rated more neighboring items than that, only the most similar ones are used.

The user's ratings are normalized by subtracting the item's mean rating from each rating prior to averaging (this is necessary to get good results with the item-mean normalization above). The item mean ratings are obtained from the model class. The resulting score function is as follows, where , the similarity between the two items:

Basket Recommendation

The item-item similarity matrix isn't just useful for generating personalized recommendations. It is also useful for ‘find similar items’ features.

The LensKit ItemBasedItemScorer and ItemBasedItemRecommender interfaces provide this functionality. ItemBasedItemScorer is like ItemScorer, except that it scores items with respect to a set of items rather than a user.

The item-based item scorer receives a basket (the set of reference items) and items (the set of items to score) vector, similar to ItemScorer. For our implementation, each item is scored with the sum of its similarity to each of the reference items in the basket.

Example Output

Predictions

Command:

./gradlew predict -PuserId=320 -PitemIds=153,260,527,588

Output:

predictions for user 320:
  153 (Batman Forever (1995)): 2.476
  260 (Star Wars: Episode IV - A New Hope (1977)): 4.262
  527 (Schindler's List (1993)): 4.167
  588 (Aladdin (1992)): 3.565

Recommendations

Command:

./gradlew recommend -PuserId=320

Output:

recommendations for user 320:
  7502 (Band of Brothers (2001)): 4.484
  1224 (Henry V (1989)): 4.423
  858 (Godfather, The (1972)): 4.408
  318 (Shawshank Redemption, The (1994)): 4.403
  1203 (12 Angry Men (1957)): 4.386
  3462 (Modern Times (1936)): 4.379
  99114 (Django Unchained (2012)): 4.376
  4973 (Amelie (Fabuleux destin d'Am?lie Poulain, Le) (2001)): 4.376
  898 (Philadelphia Story, The (1940)): 4.371
  922 (Sunset Blvd. (a.k.a. Sunset Boulevard) (1950)): 4.357

Similar Items

Command:

./gradlew itemBasedRecommend -PitemIds=153,260,527,588

Output:

1196 (Star Wars: Episode V - The Empire Strikes Back (1980)): 1.103
1210 (Star Wars: Episode VI - Return of the Jedi (1983)): 1.099
364 (Lion King, The (1994)): 1.012
595 (Beauty and the Beast (1991)): 1.005
1 (Toy Story (1995)): 0.925
500 (Mrs. Doubtfire (1993)): 0.893
5349 (Spider-Man (2002)): 0.891
480 (Jurassic Park (1993)): 0.888
1291 (Indiana Jones and the Last Crusade (1989)): 0.885
150 (Apollo 13 (1995)): 0.871