An Efficient Adaptive Attention Neural Network for Social Recommendation

Traditional recommendation algorithms based on collaborative filtering suffer from a data sparsity problem. The emergence of online social network has enriched the user's information, realizing a new way to solve recommendation tasks. Social-aware recommendation algorithms can effectively alleviate the data sparsity problem and improve the performance of recommendation systems. Despite the success of these algorithms, they have some common limitations. Most algorithms assume that social networks are homogeneous, with similar preferences among connected users. However, users may only share similar preferences in some aspects. Besides, different friends affect the user's preference in different levels. And this influence of friends on users' preference should be adaptive. Even close friends may have different influences in different decision-making processes. For example, a user may trust a friend in 'travel' but distrust this friend in 'music' because this friend had more travel experiences. Motivated by the above limitations, we designed a neural network model called adaptive attention neural network for social recommendation (ANSR) to study the interaction between a user and his or her social friends as well as infer the complex influence of the user's social relationships on the user's preferences. By utilizing the co-attention mechanism, we can not only extract the user's special attention to certain aspects of their friends but also determine the adaptive influences of different friends on the user. When the user interacts with different items, different attention weights will be assigned to the user and his or her friends, respectively. In addition, we also utilize network embedding to learn more efficient features of each user and incorporate these features into the ANSR to enhance the recommendation results. Moreover, we also conduct extensive experiments on four different real-world datasets and demonstrate that our proposed method performs better on all datasets compared with the state-of-the-art baseline methods.