Pairwise learning algorithms are a vital technique for personalized ranking with implicit feedback. They usually assume that each user is more interested in items which have been selected by the user than remaining ones. This pairwise assumption usually derives massive training pairs. To deal with such large-scale training data, the learning algorithms are usually based on stochastic gradient descent with uniformly drawn pairs. However, the uniformly sampling strategy often results in slow convergence. In this paper, we first uncover the reasons of slow convergence. Then, we associate contents of entities with characteristics of data sets to develop an adaptive item sampler for drawing informative training data. In this end, to devise a robust personalized ranking method, we accordingly embed our sampler into Bayesian Personalized Ranking (BPR) framework, and further propose a Content-aware and Adaptive Bayesian Personalized Ranking (CA-BPR) method, which can model both contents and implicit feedbacks in a unified learning process. The experimental results show that, our adaptive item sampler indeed can speed up BPR learning and CA-BPR definitively outperforms the state-of-the-art methods in personalized ranking.

Pairwise Learning; Personalized Ranking; Adaptive Sampling

--