Distributed Computing LPD

Recommender Projects

Overview

The steady growth in the number of online users has led to the emergence of various online services such as Social Networks (Google+, Facebook, Twitter), e-commerce services (Movies: IMDB, Music: last.fm, Books: Goodreads). These online services leverage personalization schemes mostly Collaborative Filtering. Collaborative filtering schemes leverage profiles of other users to improve personalization quality. On the other hand, it opens up scalability and privacy issues. Additionally, recommenders also suffer from lack of explicit feedback (cold-start) from users.

Scalable Recommender

Scalability for recommenders stems from the fact that these services need to provide personalized recommendations to millions of customers in real-time. A project here for instance would consist in experimenting scalable solutions to recommend appropriate items to web users based on some collaborative filtering protocol.
Related papers:
[1] HyRec: Leveraging Browsers for Scalable Recommenders
[2] StreamRec: A Real-Time Recommender System

Privacy-preserving Recommender

Recent research shows that customers stop using the services if they face privacy concerns. Hence, designing privacy preserving recommender is one of the major challenges at present. A project here for instance would consist in designing mechanisms which protect privacy of users in online recommender systems.
Related papers:
[1] D2P: Distance-Based Differential Privacy in Recommenders
[2] Differential privacy for neighborhood-based Collaborative Filtering

Cross-domain Recommender

Due to sparsity of data in one domain, we need to explore multiple domains to improve the user experience. A project here for instance would consist in designing an efficient mechanism to extract user profiles from one domain (e.g. Movies). These profiles should be general enough to improve recommendation quality in a different domain (e.g. Books).
Related papers:
[1] Can Movies and Books Collaborate? Cross-Domain Collaborative Filtering for Sparsity Reduction
[2] Cross-domain Recommendations without Overlapping Data: Myth or Reality?

Revenue Maximization in Recommender

The major concern for e-commerce sites is Revenue. A project here would consist in designing an efficient revenue-maximizing algorithm and demonstrate its superiority when compared to standard ones without any revenue optimizations.
Related papers:
[1] Show me the money: dynamic recommendations for revenue maximization
[2] RecMax: Exploiting Recommender Systems for Fun and Profit

Implicit Recommender

Users normally do not prefer giving explicit feedbacks like ratings. Even the ratings provided by users can vary based on their moods at any given point of time. Hence, recommenders can rely on implicit behaviour of users like clicks or consumption order. A project here for instance would consist in designing an implicit feedback (like timestamp of item consumption) based mechanism for providing recommendations to users without impacting quality significantly when compared to standard ones leveraging explicit feedback.
Related paper:
[1] A time-based approach to effective recommender systems using implicit feedback

Preference change in Recommender

Predicting preference change in recommender is an interesting research direction which can lead to better recommendation quality. For e.g. if a recommender can predict that the future user preference will vary at a given point of time with high probability then it can adapt its recommendations which standard recommenders can't. A project here for instance would consist of designing an algorithm that can predict such preference changes of users with high probability and demonstrate its superiority when compared to standard ones.
Related paper:
[1] SemanticSVD++: Incorporating Semantic Taste Evolution for Predicting Ratings

Further details of on-going works can be found at: Google Web-Alter-Egos

Contact: Rhicheek Patra