Probabilistic Topic Modeling, Reinforcement Learning, and Crowdsourcing for Personalized Recommendations

We put forward an innovative use of probabilistic topic modeling (PTM) intertwined with reinforcement learning (RL), to provide personalized recommendations. Specifically, we model items under recommendation as mixtures of latent topics following a distribution with Dirichlet priors; this can be achieved via the exploitation of crowd-sourced information for each item. Similarly, we model the user herself as an "evolving" document represented by its respective mixture of latent topics. The user's topic distribution is appropriately updated each time she consumes an item. Recommendations are subsequently based on the divergence between the topic distributions of the user and available items. However, to tackle the exploration versus exploitation dilemma, we apply RL to vary the user's topic distribution update rate. Our method is immune to the notorious "cold start" problem, and it can effectively cope with changing user preferences. Moreover, it is shown to be competitive against state-of-the-art algorithms, outperforming them in terms of sequential performance.