Adversarial Learning to Compare

Abstract

Recommendation systems tend to suffer severely from the sparse training data. A large portion of users and items usually have a very limited number of training instances. The data sparsity issue prevents us from accurately understanding users' preferences and items' characteristics and jeopardize the recommendation performance eventually. In addition, models, trained with sparse data, lack abundant training supports and tend to be vulnerable to adversarial perturbations, which implies possibly large errors in generalization.In this work, we investigate the recommendation task in the context of prospective customer recommendation in location based social networks. To comprehensively utilize the training data, we explicitly learn to compare users' historical check-in businesses utilizing self-attention mechanisms. To enhance the robustness of a recommender system and improve its generalization performance, we perform adversarial training. Adversarial perturbations are dynamically constructed during training and models are trained to be tolerant of such nuisance perturbations. In a nutshell, we introduce a Self-Attentive prospective Customer RecommendAtion framework, SACRA, which learns to recommend by making comparisons among users' historical check-ins with adversarial training. To evaluate the proposed model, we conduct a series of experiments to extensively compare with 12 existing methods using two realworld datasets. The results demonstrate that SACRA significantly outperforms all baselines.