Neural Autoregressive Collaborative Filtering for Implicit Feedback

This paper proposes implicit CF-NADE, a neural autoregressive model for collaborative filtering tasks using implicit feedback( e.g. click/watch/browse behaviors). We first convert a user's implicit feedback into a "like" vector and a confidence vector, and then model the probability of the "like" vector, weighted by the confidence vector. The training objective of implicit CF-NADE is to maximize a weighted negative log-likelihood. We test the performance of implicit CF-NADE on a dataset collected from a popular digital TV streaming service. More specifically, in the experiments, we describe how to convert watch counts into implicit "relative rating", and feed into implicit CF-NADE. Then we compare the performance of implicit CF-NADE model with the popular implicit matrix factorization approach. Experimental results show that implicit CF-NADE significantly outperforms the baseline.