際際滷

際際滷Share a Scribd company logo

Using Social- and Pseudo-Social Networks to Improve Recommendation Quality

Alan Said
Alan Said

Short paper presentation at the workshop on Intelligent Techniques from Web Personalization (ITWP2011) at the International Joint Conference on Artificial Intelligence - IJCAI-11, IJCAI2011

Convert to study materialsBETA

Transform any presentation into ready-made study materialselect from outputs like summaries, definitions, and practice questions.

1 of 13
Downloaded 11 times
1 ITWP2011, Barcelona Using Social- and Pseudo- Social Networks to Improve Recommendation Quality Alan Said, Ernesto W. De Luca, Sahin Albayrak
2 Abstract The accumulated amount of data in the digital universe reached 1.2 Zettabytes (1 billion terabytes) in 2010. 50% increase since 2008. Websites increasingly accumulate a wider variety of data on their users Without necessarily using it This paper: how can this data be used to improve recommendation
3 Outline Introduction Recommender Systems Problem statement Dataset Statistics Social and Pseudo-Social networks Approach Results
4 Introduction IMDb, one of the first online recommender systems, turned 20 on October 17th 2010. Ever since their beginning, recommender systems have, through relatively simple techniques, produced recommendations for their users Todays online systems contain more information about their users, we should use that information. Which information is important?
5 The Problem What to do with the heaps of information available? What and how to use in order to improve, or learn how to improve recommendations How should we treat Friendships? Comments? Idols? common interests? How important are these in terms of recommendation quality?
6 Dataset From the movie domain Moviepilot.de Germanys largest movie recommendation community 1M+ users 13M ratings 50K movies Subset used here 10, 000 randomly selected users with minimum 30 ratings 1.5M ratings 50, 000 comments 4, 000 friendships 170, 000 idols 25, 000 diggs
7 Social- and Pseudo-Social networks Social networks Explicit statements of friendship between users Pseudo social networks Users commenting on the same movie Users being fans of the same people Users digging the same news articles, trailers, etc. 38% of ratings performed by users with friends 45% of ratings performed by users with comments 77% of ratings performed by users who are fans 29% of ratings performed by users who digg
8 The Approach Augmentig k-Nearest Neighbor neighborhoods by using information from (pseudo) social networks Using standard Pearson Similarity Increasing the similarity of users in the same networks in order to add them to the neighborhood
9 The Approach Standard neighborhood Augmented neighborhood
10 Motivation Similarity metrics (Pearson, Jaccard, etc) are based on co- ratings Popular items often heighten similarities without adding value e.g. movies like The Matrix and The Lord of The Rings often have similar (high) ratings, even if users do not share taste Adding importance to users who share other interests filters out some of the effects of popular items.
11 Results 10 9 8 7 6 5 MAP P@10 4 3 2 1 0 Friendships Comments Fans Diggs
12 Conclusion Social and interaction (co-commenting, etc) networks seem to hold more information than standard CF is able to identify Similarity metrics do not always tell the complete truth ToDos: Find items that are important for establishing similarity between users Investigate what other information can be used for measuring similarities
13 Questions? Thank you!

More Related Content

Using Social- and Pseudo-Social Networks to Improve Recommendation Quality

  • 1. 1 ITWP2011, Barcelona Using Social- and Pseudo- Social Networks to Improve Recommendation Quality Alan Said, Ernesto W. De Luca, Sahin Albayrak
  • 2. 2 Abstract The accumulated amount of data in the digital universe reached 1.2 Zettabytes (1 billion terabytes) in 2010. 50% increase since 2008. Websites increasingly accumulate a wider variety of data on their users Without necessarily using it This paper: how can this data be used to improve recommendation
  • 3. 3 Outline Introduction Recommender Systems Problem statement Dataset Statistics Social and Pseudo-Social networks Approach Results
  • 4. 4 Introduction IMDb, one of the first online recommender systems, turned 20 on October 17th 2010. Ever since their beginning, recommender systems have, through relatively simple techniques, produced recommendations for their users Todays online systems contain more information about their users, we should use that information. Which information is important?
  • 5. 5 The Problem What to do with the heaps of information available? What and how to use in order to improve, or learn how to improve recommendations How should we treat Friendships? Comments? Idols? common interests? How important are these in terms of recommendation quality?
  • 6. 6 Dataset From the movie domain Moviepilot.de Germanys largest movie recommendation community 1M+ users 13M ratings 50K movies Subset used here 10, 000 randomly selected users with minimum 30 ratings 1.5M ratings 50, 000 comments 4, 000 friendships 170, 000 idols 25, 000 diggs
  • 7. 7 Social- and Pseudo-Social networks Social networks Explicit statements of friendship between users Pseudo social networks Users commenting on the same movie Users being fans of the same people Users digging the same news articles, trailers, etc. 38% of ratings performed by users with friends 45% of ratings performed by users with comments 77% of ratings performed by users who are fans 29% of ratings performed by users who digg
  • 8. 8 The Approach Augmentig k-Nearest Neighbor neighborhoods by using information from (pseudo) social networks Using standard Pearson Similarity Increasing the similarity of users in the same networks in order to add them to the neighborhood
  • 9. 9 The Approach Standard neighborhood Augmented neighborhood
  • 10. 10 Motivation Similarity metrics (Pearson, Jaccard, etc) are based on co- ratings Popular items often heighten similarities without adding value e.g. movies like The Matrix and The Lord of The Rings often have similar (high) ratings, even if users do not share taste Adding importance to users who share other interests filters out some of the effects of popular items.
  • 11. 11 Results 10 9 8 7 6 5 MAP P@10 4 3 2 1 0 Friendships Comments Fans Diggs
  • 12. 12 Conclusion Social and interaction (co-commenting, etc) networks seem to hold more information than standard CF is able to identify Similarity metrics do not always tell the complete truth ToDos: Find items that are important for establishing similarity between users Investigate what other information can be used for measuring similarities
  • 13. 13 Questions? Thank you!