In e-commerce systems, user preference can be inferred from multivariate implicit feedback (i.e., actions). However, most methods merely focus on homogeneous implicit feedback (i.e., purchase). In this paper, we adopt another two typical actions, i.e., view and like, as auxiliaries to enhance purchase recommendation, whereby a trinity Bayesian personalized ranking (TBPR) method is proposed. Specifically, we introduce trinity preference to investigate the difference of users' preference among three types of items: 1) items with purchase action; 2) items with only auxiliary actions; 3) items without any action. Empirical study on the real-world dataset demonstrates that our method significantly outperforms state-of-the-art algorithms.