MSc IS Thesis Proposal by NGUYEN Tiep Trong

Community-Based Collaborative Learning for Personalized Recommendation

With the rapid growth of the Web and social media, understanding users' preferences is a key task in recommender systems. The aim of this thesis is to exploit the role of "community" in mining users' behaviors. Our work draws inspiration from collaborative filtering, where users share their preferences through similar ratings; however, we seek to explore the effects of community rather than individual links.

In the first part, we seek to uncover "latent" communities users belongs to based on their similar behaviors. Within the framework of reinforcement learning, users are dynamically clustered via their feedbacks from system interactions.

In the second part, we deal with cold-start users based on the "social" community, with the goal of finding the right structure for social integration under undirected graphical models like restricted Boltzmann machine and deep Boltzmann machine.

Lastly, we remedy both cold-start users and items into an integrated model, where topic- and community-distributions are bridged into a common latent space. The proposed method indicates promising results, especially in sparse data settings.