Collaborative filtering (CF)-based recommendation systems that rely on user-item history interactions often suffer from the data sparsity problem. Social-based recommendation methods have become one of the successful methods to address this problem. However, few works have focused on the sparsity problem of social data. As real-world social networks are usually sparse, the observed relationships in social networks can only represent a limited part of a person's real social network. The sparse social data will degrade the performance of the existing social-based algorithms. Also, the influence of a user's friends on their friends is dynamic: even the same friend may impact the target user in different decision-making processes. It is difficult for an the end-to-end deep learning-based model to provide underlying reasons for the recommendation results. To this end, we propose a novel deep learning-based model to extract useful missing links as auxiliary social information to enrich the users' features for item recommendation. The framework is composed of two major components: a missing links identifier module that generates useful social links from a heterogeneous information network to enrich the user's social profile and enhance the social-based recommendation model, and an attention-based recommendation module that assigns different scores for each friend with regard to different candidate items to adaptively evaluate the quality of different social links. An attention-based fusion strategy is proposed to improve the interpretability of the recommendation system by assigning non-uniform weight to different factors. Extensive experiments on three published datasets show that our proposed method achieves better performance than other state-of-the-art methods.