It Takes Variety to Make a World

1
It Takes Variety toMake a World

• Diversification in Recommendation Systems
• Cong Yu1, Laks Lakshmanan2, Sihem Amer-Yahia1
• 1Yahoo! Research NYC 2University of British
  Columbia
• March 25th 2009 _at_ EDBT

2
Recommendation An Increasingly Important and
Ubiquitous Paradigm on the Web

• The Recommendation Paradigm
• Suggest content (in most cases, items) to users
  based on her profile and past activities.
• Why Recommendation?
• Search queries can be generic e.g., gt90 of
  Yahoo! Travel queries are general descriptions
  like family trip.
• More so for Social Content Sites ...

3
Recommendations on Social Content Sites

• Social Content Sites
• Sites where users make friends and share contents
• E.g., del.icio.us, Flickr, etc.
• Recommendation is an indispensible information
  exploration paradigm on social content sites.
• The rich activities and user connections provide
  lots of opportunities for generating
  recommendations.

4
Challenges in Recommendation

• While relevance is important, other factors are
  critical too
• Diversity avoid returning items that are too
  similar to each other.
• Novelty avoid returning items that users are
  likely to know already.
• Serendipity aim to return less relevant items
  that might give users a pleasant surprise.
• Result Diversification

From the pool of relevant items, identify a list
of items that are dissimilar to each other and
maintain a high cumulative relevance, i.e.,
strike a good balance between relevance and
diversity.
5
Existing Solutions

• Attribute-Based Diversification
• Follow Three Steps
• Obtain attributes of each relevant items
• Define a pair-wise item-to-item distance function
  based on those attributes
• Perform Diversification
• Optimizing an overall score as a weighted
  combination of relevance and distance
• Constraining either relevance or distance,
  maximizing the other

6
Problems with ExistingAttribute-Based
Diversification

• Lack of attributes for objects
• URLs in del.icio.us and photos in Flickr
• Overhead for retrieving attributes for certain
  recommendation strategies
• Difficulties in estimating the correct
  parameters/thresholds for diversification
  algorithms

Our Solutions Explanation-Based
Diversification Dynamic Diversification
Algorithms
7
Main Contributions

• Formalized the Notion of Explanation-Based
  Diversification
• Designed and Implemented Algorithms for
• Scalable Similarity Computation
• Explanation Generation
• Diversification
• Experimentally Evaluated
• The characteristics of diversifications
  algorithms
• The practicality of explanation-based
  diversification
• The performance overhead of explanation-based
  diversification

8
Outline

• Motivation
• Problem Definition
• Algorithms
• Similarity Computation
• Recommendation Generation with Explanation
• Diversification
• Experimental Evaluation
• Conclusion

9
Recommendation Strategies Overview

• Item-Based Strategies
• Estimate the rating of an unrated item (i) by the
  user (u) based on its similarity to items already
  rated and how u rated those items.
• Collaborative Filtering Strategies
• Estimate the rating of i by u based on how us
  similarity network (either explicit or implicit)
  rated i.

10
Explanation

• Basic Notion
• The set of objects because of which a particular
  item is recommended to the user
• Explanation for Item-Based Strategies
• Explanation for Collaborative Filtering Strategies

11
Explanation-Based Diversity

• Pair-wise diversity distance between two
  recommended items
• Standard similarity measures like Jaccard
  similarity and cosine similarity
• E.g. (Distance based on Jaccard similarity)
• Diversity for the set of recommended items (S)

12
Benefits of Explanation-Based Diversification

• Applicable to items without attributes or whose
  attributes are difficult to analyze
• Common on social content sites
• Explanations are by-products of many
  recommendation processes
• They can be maintained with little overhead

13
Top-K Recommendation with Diversification Given
a user u, find a subset S from the set of
candidate items, such that S k and the
overall relevance of items in S and the diversity
of S are balanced.
14
Outline

• Motivation
• Problem Definition
• Algorithms
• Similarity Computation (briefly)
• Recommendation Generation with Explanation
  (briefly)
• Diversification
• Experimental Evaluation
• Conclusion

15
Diversification Balance between Relevance and
Diversity

• Relevance cumulative relevance of all items in
  the result set
• Diversity average distance of all item pairs in
  the result set as described earlier
• Ideal Scenario identify a top-k result set that
  maximizes both relevance and diversity
• Such top-k set is often impossible to find

16
Naïve Solution Maximize Relevance
17
Naïve Solution Maximize Diversity
18
Smarter Solutions

• Eliminate items with scores below a threshold and
  choose k items among the remaining with the
  maximum diversity
• Eliminate item pairs with distance below a
  threshold (by removing the item with lower score)
  and choose k items with highest scores

Algorithm Swap combines both!
19
Algorithm Swap

• Sort candidate items according to their relevance
• Start by adding the K most relevant items to the
  result set
• Go through the rest of the candidate one by one,
  swap an item into the result set if the item
• Increases the set diversity above a certain
  threshold
• Does not drop the relevance by a certain
  threshold
• A simple top-2 example

c1
Results
Results
Results
c1
c1
c2
c1
c3
c3
c4
c3
c3
relevance c1 0.9 c2 0.6 c3 0.4 c4 0.3
c2
c4
initial set
diversity increase out-weighs the relevance drop
no change
20
Challenge The appropriate thresholds are often
difficult to identify to produce the
right top-K recommendations
21
Algorithm Iterative Greedy

• Dynamically identify the thresholds
• Establish two diversity bounds, Upper and Lower
  Bounds.
• At each iteration, scan the candidates
• items passing upper bound go to KeepList
• items not passing lower bound go to DiscardList.
• At the end of each iteration, bounds are adjusted
• Stop when exactly K items are generated

22
Algorithm Iterative Greedy, contd
If DivList KeepList lt K
DivList
KeepList
pass
pass
UB
UB B
Candidates
B (UBLB)/2
Next Iteration
Candidates
LB B
LB
no-pass
no-pass
SimList
DiscardList
23
Brief Overview of Similarity Computation

• Explicit network is not enough
• E.g., only 10 of users have at least one friend
  in del.icio.us
• Similarities between users can be generated based
  on their activities
• Costly with pair-wise comparison
• E.g., 1 million users gt 1 trillion comparisons
• Only a small fraction of those comparisons result
  in similarity above a given threshold
• Item-Based Similarity Computation
• Organize items based on the number of raters
• Start with items with the largest number of
  raters
• Compare two users only if they share enough rated
  items
• Details in the paper

24
Brief Description of Recommendation Generation
with Explanation

• Post-Processing Approach
• Generate the recommendation result set
• Generate candidates
• Compute scores for candidates
• Sort the candidates
• For each item in the result set, fetch its
  explanations
• Integrated Approach
• Maintain the list of similar items or similar
  users when the candidates are being generated
• Create the explanations for each item while the
  scores are being computed
• Details in the paper

25
Outline

• Motivation
• Problem Definition
• Algorithms
• Similarity Computation (briefly)
• Recommendation Generation with Explanation
  (briefly)
• Diversification
• Experimental Evaluation
• Conclusion

26
Experimental Data

• Real world data sets
• del.icio.us online bookmark sharing site
• Y! Movies Yahoo!s online movie sharing site

27
Result Comparison
28
Explanations can Serve as a Good Basis for
Diversification

• Leveraging the Yahoo! Movies data set, we compare
  diversified results obtained based on explanation
  with those obtained based on attributes

29
Overhead of Diversification is Small
30
Outline

• Motivation
• Problem Definition
• Algorithms
• Similarity Computation (briefly)
• Recommendation Generation with Explanation
  (briefly)
• Diversification
• Experimental Evaluation
• Conclusion

31
Conclusion

• Recommendation is becoming an indispensible
  information exploration paradigm
• Explanation-Based Diversification is a practical
  alternative to attribute-based diversification
• Algorithms Swap and Iterative Greedy strike a
  good balance between relevance and diversity
• Questions?