Collaborative Topic Model for Poisson distributed ratings

We present Collaborative Topic Model for Poisson distributed ratings (CTMP), a hybrid and interpretable probabilistic content-based collaborative filtering model for recommender system. The model enables both content representation by admixture topic modelling, and computational efficiency from Poisson factorization living together under one tightly coupled probabilistic model, thus addressing the limitation of previous methods. CTMP excels in predictive performance under different real-world recommendation contexts, and easily scales to big datasets, while recovering interpretable user profiles. Moreover, our empirical study also shows strong evidence that sparsity in the estimates of topic mixture can be recovered via learning, despite not being specified in the model. The sparse representation derived from CTMP would allow efficient storage of the item contents, consequently providing a computational advantage for other tasks in industrial settings.