Understanding users' behavior and predicting their future purchase are critical for e-commerce companies to boost their revenue. While explicit user feedback such as ratings plays the most significant role in eliciting users' preferences, such feedback is scarce, which prompts the need for leveraging more abundant implicit user feedback such as purchase record. Consequently, recent studies focused on leveraging users' past purchase record to predict their purchase. However, their performance is not satisfactory due to 1) the lack of purchase history of users, and 2) more importantly the ill-posed assumption of non-purchased items equally being considered as negative feedback. In this paper, we define new pairwise relationships among items aiming at overcoming the limitations of existing works, and propose a novel method called P3S that stands for modeling pairwise relationships among three disjoint item sets, which leverages users' click record in conjunction with their purchase record. Precisely, we partition the items into three disjoint sets based on users' purchase and click record, define new pairwise relationships among them with respect to users, and reflect these relationships into our pairwise learning-to-rank method. Experiments on two real-world datasets demonstrate that our proposed method outperforms the state-of-the-art baselines in the task of predicting users' future purchase in e-commerce.