New Probabilistic Models for Recommender Systems with Rich Contextual and Content Information

Abstract

This project is focused on the design of probabilistic models for recommender systems and collaborative ltering by extending and creating new models to include rich contextual and content information (content, user social network, location, time, user intent, etc), and developing scalable approximate inference algorithms for these models. The working hypothesis is that big data analytics combined with probabilistic modelling, through automatically mining of various data sources and combining di erent latent factors explaining the user interaction with the items, can be used to better infer the user behaviour and generate improved recommendations. Fundamentally we are interested in the following questions: 1) Does additional contextual information improve the quality of recommender systems? 2) What factors (features, model, methods) are relevant in the design of personalized systems? 3) What is the relation between the social network structure, the user model and the information need of the user? How does the social context interferes with user preferences? How the evolution of the social network structure can explain changes in the user preference model? 4) Does the choice of approximate inference method have a signi cant impact on the quality of the system (quality- efficiency trade-offs)? To address some of this questions we started by proposing a model (Figure 1) based on Poisson factorization models [2], combining a social factorization model [1] and a topic based factorization [3]. The main idea is to combine content latent factor (topic, tags, etc) and trust between users (trust weight in a social graph) in a way that both sources of information have additive e ects in the observed ratings. In the case of Poisson models, this additive constraint will induce non-negative latent factors to be more sparse and avoid overfitting (in comparison the Gausian based models [2]. The main objective at this point is to compare models that incorporated both source of information (content and social networks). The next steps will include empirical validation. Concluding, we are interested in the interplay between large scale data mining and probabilistic modeling in the design of recommender systems. One initial approach we are pursuing is to model content and social network feature in a Poisson latent variable model. Our main objective in the future is the development of methods with competitive computational complexity to perform inference using het- erogeneous data in dynamical probabilistic models, as well as exploring the scalability limits of the models we propose.