Self-Adaptive Deep Asymmetric Network for Imbalanced Recommendation

Recently, rating prediction based on deep models has been a promising way to the personalized recommendation for the problem of information overloaded, which is a crucial problem to evaluate users' preferences for different items. In reality, the rating distribution in training data being mostly imbalanced in that the number of low ratings is often much less than that of high ratings, which is the biased prediction or imbalanced dataset problem in the recommendation. However, due to the degree of imbalance always varying greatly in different scenarios, most of the existing imbalanced recommendation methods fix the model parameters and cannot adapt to the change of the ratio between frequent and rare ratings in different datasets. In this article, we propose a novel Self-adaptive Deep Asymmetric Network (SDAN) to address the problem of imbalanced recommendation, which can automatically adjust the model based on the degree of imbalance in the recommendation dataset and accurately predict users' frequent and rare ratings of items. Specifically, two different autoencoders are first introduced to learn the user-based and item-based feature representations. Then the self-adaptive deep network is designed to automatically emphasize more on rare ratings and compensate for the bias prediction problem. Finally, the reconstructed approximated matrix with the features of users and items is utilized for predicting the ratings of user-item pairs. Through comprehensive experiments on several public recommendation datasets with varying imbalanced degrees, our SDAN consistently outperforms state-of-the-art methods.