3. INTRODUCTION (1/3)
• Recommender system suggest items of interest to users.
• Collaborative filtering (CF) users rating information to recommend
items based on similarity.
• The drawback: more appropriate for static settings.
• In real world data, the new users and items should be incorporated
into model recommendations in an online manner. The
incremental CF can handle the need.
3
4. INTRODUCTION (2/3)
• A few published approaches of the incremental CF:
• Sarwar et al. proposed an online CF strategy using singular value
decomposition, SVD.
• Das et al. proposed a scalable online CF using MinHash
clustering, PLSI and co-visitation counts.
• In K-NN, similarity parameters such as correlation can be
updated incrementally during online phase.
• George and Merugu used Bregman co-clustering as a scalable
incremental CF approach for dynamic settings. (ICDM’05)
4
5. INTRODUCTION (3/3)
• This paper propose an incremental CF method that is both
scalable and accurate.
• The main contribution of this paper:
• An evolutionary Bregman co-clustering algorithm
• An ensemble strategy to give better predictions.
5
6. INCREMENTAL CF
(1/3)
• In a CF problem, there are U users and V items.
• Users have provided a number of explicit ratings for items.
• rui is the rating of user u for item i.
• There are two phases in a CF algorithm:
• Offline phase: training based on known ratings
• Online phase: unknown ratings are estimated using the output of offline
phase.
• In incremental CF, the data available during online phase is
incorporated into future predictions.
6
7. BASELINE
ALGORITHM
• The simplest way to predict a rating is the global average r- of all
ratings.
• However, some users tend to rate higher and some items are more
popular. Including user bias and item bias in rating, the prediction is
given by:
• r-u: the average ratings by user u.
• r-i: the average of ratings for item i.
• nu: the number of ratings for user u.
• ni: the number of ratings for item i.
• Snu,w and Sni, w are the support function for user u and item i.
7
8. INCREMENTAL CF VIA CO-
CLUSTERING (ICDM’05) (1/2)
• Clustering refers to partitioning similar objects into groups, while co-
clustering partitions two different kinds of objects simultaneously.
• As suggested in George’s paper, the prediction is as follows:
• where k=(u) is the user cluster assigned to user u.
• l=(i) is the item cluster assigned to item i.
• r-kl is the average of ratings belonging to users in user cluster k
and items in item cluster l.
• (r-u-r-k) is the bias of user u.
• (r-i-r-l) is the bias of item i.
8
9. INCREMENTAL CF VIA CO-
CLUSTERING (ICDM’05) (2/2)
• George used the Bregman co-clustering algorithm, which has two
phases, updating user clusters and updating item clusters, to produce
the co-cluster results.
• In the online phase, the prediction is as follows:
• Incremental training is achieved by using new ratings to update the
average parameters (r-kl, r-u, r-k, r-i, r-l).
• However, new users or items are not assigned to clusters during the
online phase.
9
10. INCREMENTAL EVOLUTIONARY
CO-CLUSTERING (1/4)
• If the support Sv,w (number of available ratings) for a user or items is
low, the co-clustering approach will not provide good predictions for
them.
• As a strategy, users and items with low support are removed from the
training phase so that training is both more effective and efficient.
• The drawback of Eq. (3),
• It incorporates (r-kl, r-k, r-l) from co-clustering solution that is not
necessarily reliable. (r-k and r-l is close to r-)
• Using only the block average r-kl for prediction ignores user and
item bias which results in poor accuracy as well.
10
11. INCREMENTAL EVOLUTIONARY
CO-CLUSTERING (2/4)
• The revised rating prediction with co-clustering residuals is model as
• Eq.(5) is come from the support function of Eq. (1) set to 1.
• The ui is the correction parameter for (1).
• For known rating, Eq. (5) can be rewritten as
• ui can be interpreted as the residual of the prediction via (1).
• For implementing co-clustering, it is enough to work with the following objective
function.
• Where wui is 1 if rating rui exists in training data and otherwise is 0.
• (u)(i): the block average of residuals for user cluster (u) and item cluster (i).
11
12. INCREMENTAL EVOLUTIONARY
CO-CLUSTERING (3/4)
• The prediction strategy of old user - old item,
• and otherwise
• The ensembles are used to improve the accuracy of a method using a group
of predictors, while increasing the running time linearly with the number of
ensemble elements.
• Let p denote a co-clustering solution and P be the number of co-clustering
solutions we use in the model. We can predict with
• zulp is the average error of prediction for user u and item cluster l in p.
• zikp is the average error of prediction for item i and user cluster k in p.
12
13. INCREMENTAL EVOLUTIONARY
CO-CLUSTERING (4/4)
• In this paper, it is trivial to find an appropriate cluster for a new user or new
item.
• Let u be a new user who has provided some ratings.
• If a sufficient number of rated items exits in the current co-clustering
solution (sub-matrix), then the new user’s cluster can be found using
• nuh is the number of times user u has rated the items belonging to item
cluster h during the online phase.
• -uh is the average of residuals for those ratings.
• g: user cluster
• A similar procedure finds the cluster of a new item.
13
14. INCREMENTAL TRAINING
ALGORITHM
• numberIn() is the number of
ratings a user u (item i) has in
the co-clustering solution
which is defined by hnuh
(gnig).
Trust the information
for incorporating new
user or new item.
• The new users and items will
not receive any prediction,
those are predicted by Eq. (1).
14
15. EVOLUTIONARY ALGORITHM A group of co-clustering solutions is
randomly generated and locally
• A population-based search optimized via Bregman co-clustering.
approach
• Goal: find better solutions by
combining the current
solutions.
• Every evolutionary algorithm
has three main step
• Selection
• Crossover
• Replacement
Worst solution
15
16. CROSSOVER
ALGORITHM
• Let X be a NK assignment
matrix
• An element x=(u, k) is 1, if object
u is assigned to cluster k and 0
otherwise.
• qr is the intersection between
cluster q and cluster r.
• (k) is the largest intersection.
16
18. EXPERIMENTAL
RESULTS (1/3)
• The experiment dataset: Movielens dataset consisting of 100,000 ratings
(1-5) by 943 users on 1682 movies.
• Evaluation metrics: Mean Absolute Error (MAE)
• Comparison methods:
• Baseline
• COCL: George, ICDM’05
• ECOCL: Evolutionary co-clustering without ensembles
• ECOCLE: Evolutionary co-clustering with ensembles.
• IKNN: Incremental KNN method.
• SVD
18
19. EXPERIMENTAL
RESULTS (2/3)
• The experiment use the 5-fold
cv. to get average MAE.
• The incremental training based
on three different strategies.
• “20%-80%”: 20% of data was
used for offline training, 80%
for incremental training.
19
20. EXPERIMENTAL
RESULTS (3/3)
• The offline phase of ECOCLE
needs more time due to the
evolutionary algorithm.
• Online time is the sum of both
incremental training and
prediction.
• ECOCL and IKNN have
similar online speeds, while
the accuracy for ECOCLE is
much higher.
20
21. CONCLUSION
• Online CF methods that can incorporate new data in real time are
advantageous in many practical situations.
• However, this problem has not been adequately addressed.
• This paper extended the idea of CF via co-clustering to satisfy this need.
• The empirical results showed the proposed ECOCLE avchived very good
accuracy compared to other incremental methods.
• Training time was comparatively slow, but still manageable.
21