Modeling Implicit Communities in Recommender Systems

Abstract

In recommender systems, a group of users may have similar preferences on a set of items. As the groups of users and items are not explicitly given, these similar-preferences groups are called implicit communities (where users inside same communities may not necessarily know each other).

Implicit communities can be detected with users’ rating behaviors. In this paper, we propose a unified model to discover the implicit communities with rating behaviors from recommender systems.

Following the spirit of Latent Factor Model, we design a bayesian probabilistic graphical model which generates the implicit communities, where the latent vectors of users/items inside the same community follow the same distribution. An implicit community model is proposed based on rating behaviors and a Gibbs Sampling based algorithm is proposed for corresponding parameter inferences. To the best of our knowledge, this is the first attempt to integrate the rating information into implicit communities for recommendation.

We provide a linear model (matrix factorization based) and a non-linear model (deep neural network based) for community modeling in recsys.

Extensive experiments on seven real-world datasets have been conducted in comparison with 14 state-of-art recommendation algorithms. Statistically significant improvements verify the effectiveness of the proposed implicit community based models. They also show superior performances in cold-start scenarios, which contributes to the application of real-life recommender systems.