Neural Poisson Factorization

In this work, we focus on dealing with a sparse users' feedback matrix and short descriptions/contents of items in recommender systems. We propose the Neural Poisson factorization (NPF) model which is a hybrid of deep learning and Poisson factorization. While Poisson factorization is suitable to model discrete, massive and sparse feedback, using a deep neural network and pre-trained word embeddings can learn hidden semantic in short item descriptions well. Therefore, NPF overcomes the limitation of existing models when dealing with short texts and a sparse feedback matrix. Moreover, we develop a random view algorithm based on stochastic learning for our model, in which each user is only viewed a random subset of items and his/her feedback on the subset is used to update his/her representation in each iteration. This approach is reasonable because each user can only know or view a partial subset of items when surfing a system. Extensive experiments illustrate the significant advantages of NPF over content-based matrix factorization methods and others that ignore item descriptions.