RecSys 2011 Review
Qi Zhao11-01-2011
Outline
• Overview• Sessions– Algorithms– Recommenders and the Social Web– Multi-dimensional Recommendation, Context-
awareness and Group Recommendation– Methodological Issues, Evaluation Metrics and Tools– Human factors– Emerging Recommendation Domains
• Conclusion
Overview• Participants– Student, professor– Research Institutes, like Yahoo! Research, eBay Research,
Microsoft Research, etc– Industry. Twitter, Google, Facebook, Netflix, LinkedIn, etc
• Oral papers, posters, workshops, demos• Themes– Algorithm– Recommendation and the Social Web– Multi-Dimensional Rec, Group Rec, Context-Aware Rec– Evaluation Metric– Human factors– Emerging Domains
Session: Algorithm
• Major issues to tackle– Cold start
Generalizing Matrix Factorization Through Flexible Regression Priors
• Motivation– Warm-start scenario: low-rank factorization +
regularization– Zero-mean regularization– Handle cold-start scenario
• New users • Approach– GMF
• Regularization based on Non-linear regression on user /item feature
Shared Collaborative Filtering
• How it works?– Leverage the data from other parties to improve
own CF performance
• Issues– Privacy concerns when sharing the community
data
Session: Recommender Systems and the Social Web
Recommendation in Social Rating Networks
• Social Rating Network– User-user relationship– User express ratings over some items– Example: Epinions, Flixter,
• Why use social networks in recommendation?– Selection and social influences by sociologist– Selection: tendency to relate to people with similar attributes
• SNR: similar rating behavior– Social influence: adopting ratings from friends
• Selection and social influence drive the formation of like-minded and well-connected users.
• Challenges– Mixed groups, social relations– Generalized Stochastic Block Model
• Mixed group membership for both users and items
Personalized PageRank Vectors for Tag Recommendations: Inside FolkRank
• Setting: Folksonomy– User, Tags, Resources(flickr, del.icio.us, etc)– User assign tags to resources.
• Problem– Ranking tag, user and resource– Tag recommendation
• Main contribution– Present and formalize the FolkRank model – Present FolkRank-like model which provides fast tag
recommendation
Session: Multi-dimensional Recommendation, Context-awareness and Group
Recommendation
Multi-Criteria Service Recommendation Based on User Criteria Preference
• Using multiple criteria to value the product or service– E.g. Restaurant – price, location, quality of food,
service speed, etc
• User has her own preference over the attributes
• Cluster users based on their preference– Prediction based on users within the same cluster
The Effect of Context-Aware Recommendations on Customer Purchasing Behavior and Trust
• Content-Aware Recommendation Systems(CARS)– Additional information like location, time, your companies,
etc• Effect on Purchasing Behavior– Accuracy– Trust. Recommendation should be credible and objective.
• Methodology– Controlled experiment– Three methods: content-based, CARS, random– Metric: accuracy, diversity(entropy)– Purchasing change: Money spend on the product
Group Recommendation
• Recommendations for a group of people instead of individuals– E.g. people sitting around watching tv
• The challenge– Aggregated preference might be diverse– Depend on the group’s characterizer– Homogeneous or Heterogeneous• Similar demographic information or not
Session: Methodological Issues, Evaluation Metrics and Tools
OrdRec: An Ordinal Model for Predicting Personalized Item Rating Distribution
• Common views upon feedbacks– Numerical values– Apply Collaborative Filtering
• About numerical ratings– Different users have their own internal scale– Hard to assign a numerical value– Ranking products through comparing
• Humans are more consistent when comparing products than giving absolute scores• Ordinal
– Express relative preference over items• Evaluation
– RMSE– Fraction of Concordant Pairs(FCP)– OrdRec outperforms existing approaches: SVD++, RBM, MultiMF
Session: Human factors
A User-Centric Evaluation Framework for Recommender Systems
• ResQue(Recommender system’s Quality of User Experience)– Understanding issues of RecSys
• Evaluation Layers– Perceived system qualities – User’s belief– Subjective attitude– Behavioral intention
• Experiment Design– Survey on 239 participants
Cont.
Session: Emerging Domains
• Yahoo! Music Recommendation: Modeling Music Ratings with Temporal Dynamics and Item Taxonomy
• CrimeWalker: A recommendation Model for Suspect Investigation
• Personalized Activity Stream: Sifting through the “River of News”
Conclusion
• Modeling the Recommendation– Collaborative Filtering– Incorporating additional features
• Evaluation Metrics– Accuracy, Diversity, Novelty, etc
• Adapt to Constantly Changing Internet Ecosystem– Social Network– Realtime Activity Stream