Recommender systems aim to predict the content that a user would like based on observations of the online behaviour of its users. Research in the Information Access group addresses different aspects of this problem, varying from how to measure recommendation results, how recommender systems relate to information retrieval models, and how to build effective recommender systems (note: last Friday, we won the ACM RecSys 2013 News Recommender Systems challenge). We would like to develop a general methodology to diagnose weaknesses and strengths of recommender systems. In this talk, I discuss the initial results of an analysis of the core component of collaborative filtering recommenders: the similarity metric used to find the most similar users (neighbours) that will provide the basis for the recommendation to be made. The purpose is to shed light on the question why certain user similarity metrics have been found to perform better than others. We have studied statistics computed over the distance distribution in the neighbourhood as well as properties of the nearest neighbour graph. The features identified correlate strongly with measured prediction performance - however, we have not yet discovered how to deploy this knowledge to actually improve recommendations made.
2. Recommendation
• Informally:
– Search for information “without a query”
• Three types:
– Content-based recommendation
– Collaborative filtering (CF)
• Memory-based
• Model-based
– Hybrid approaches
3. Recommendation
• Informally:
– Search for information “without a query”
• Three types:
– Content-based recommendation
– Collaborative filtering
• Memory-based
• Model-based
– Hybrid approaches
Today’s focus!
4. Collaborative Filtering
• Collaborative filtering (originally introduced by
Patti Maes as “social information filtering”)
1. Compare user judgments
2. Recommend differences between
similar users
• Leading principle:
People’s tastes are not randomly
distributed
–A.k.a. “You are what you buy”
5. Collaborative Filtering
• Benefits over content-based approach
– Overcomes problems with finding suitable
features to represent e.g. art, music
– Serendipity
– Implicit mechanism for qualitative aspects like
style
• Problems: large groups, broad domains
15. Research Question
• How does the choice of similarity measure
determine the quality of the
recommendations?
16. Sparseness
• Too many items exist, so many ratings will
be missing
• A user’s neighborhood is likely to extend
to include “not-so-similar” users and/or
items
17. “Best” similarity?
• Consider cosine similarity vs. Pearson
similarity
• Most existing studies report Pearson
correlation to lead to superior
recommendation accuracy
18. “Best” similarity?
• Common variations to deal with sparse
observations:
– Item selection:
• Compare full profiles, or only on overlap
– Imputation:
• Impute default value for unrated items
– Filtering:
• Threshold on minimal similarity value
21. Distance Distribution
• In high dimensions, nearest neighbour is unstable:
If the distance from query point to most data points is less than
(1 + ε) times the distance from the query point to its nearest
neighbour
Beyer et al. When is “nearest neighbour” meaningful? ICDT 1999
23. Distance Distribution
• Quality q(n, f):
Fraction of users for which the similarity
function has ranked at least n percent of
the user community within a factor f of the
nearest neighbour’s similarity value (well... its
corresponding distance)
25. NNk
Graph
• Graph associated with the top k nearest
neighbours
• Analysis focusing on the binary relation of
whether a user does or does not belong to
a neighbourhood
– Ignore similarity values (already included in
the distance distribution analysis)
27. MRR vs. Features
• Quality:
– If most of the user population is far away, high
similarity correlates with effectiveness
– If most of the user population is close, high
similarity correlates with ineffectiveness
29. Conclusions (so far)
• “Similarity features” correlate with
recommendation effectiveness
– “Stability” of a metric (as defined in database
literature on k-NN search in high dimensions)
is related to its ability to discriminate between
good and bad neighbours
30. Future Work
• How to exploit this knowledge to now
improve recommendation systems?