关键词: 推荐算法, 残差, 注意力机制, 协同过滤

Abstract: With the development of Internet technology, online users usually need to spend a lot of time and effort searching the content they are interested in from the huge amount of information, so the problem with information overload arises. As the most simple and efficient information filtering technique, recommender systems (RS) can fully analyze the previous preference behaviors of users and build recommendation models to predict user preferences, and recommend items most likely to be of interest to active users, which greatly improves user satisfaction and effectively alleviates the information overload problem. Especially, recommendation models based on deep learning (DL) show good application potential owing to making full use of the low-order and high-order features contained in the data. However, DL-based recommendation models are prone to the problem with gradient disappearance when capturing the high-order features because of data sparseness and user differentiation. Meanwhile, to the best of our knowledge, few studies consider personalized user preferences when integrating the high-order and low-order features.
To address the above issues, this paper introduces the concept of residual into a DL-based recommendation model and proposes a residual neural collaborative filtering model integrating attention mechanism. Firstly, our model adopts a multi-layer neural network to effectively extract the high-order features between users and items; and a dense residual connection is used to transfer the low-order features to the hidden layer at the back end of the network, which can avoid the information loss caused by the layer-by-layer transmission of information in the traditional multi-layer perceptron (MLP). Subsequently, in consideration of the personalized recommendation needs of different users, we use an attention mechanism to adaptively integrate the low-order and high-order features to generate a set of feature vectors that can better express the preferences of different users. Finally, the preference ratings for all items are evaluated according to the feature vectors, and selecting Top-K items is recommended to target users through a sorting method.
To validate the effectiveness of the proposed model, three public datasets with different sparse levels are used for the experiments. And, four evaluation metrics, namely Precision, Recall, F1 value, and NDCG, are used to test the performance of our model. Firstly, we design ablation experiments to illustrate the advantages of the used residual channel and attention mechanism. The experimental results show that our model achieves the best results among them, which indicates that both the residual channel and attention mechanism can better capture the user interest points and improve the recommendation accuracy of the model. Then, we compare four representative methods, called MF, NCF, AFN, and Wide & Deep, to further verify the effectiveness of the proposed model. The results show that our model has the better performances compared with the comparison methods in terms of various evaluation metrics on all datasets. Compared with the closest competitor, Wide & Deep, the performance of our model will increase by at least 5% and 2.6% when the length of recommendation list is fixed to 10 in the metrics F1 value and NDCG, respectively. Thus, the proposed model effectively improves the quality of RSs and has good application value.
The main limitation of our model in this paper is that it only utilizes implicit feedback information to predict user preferences, while ignoring the influence of explicit feedback information, e.g., ratings, reviews, etc., on the degree of user preferences, making this model unable to take into account the consistency and reliability of user purchase intentions and interest preferences. Therefore, future work can be done in two aspects. On the one hand, we can effectively integrate the two types of information to fully capture the common features of users' real preferences to enhance the prediction ability and recommendation reliability of the proposed model. On the other hand, user's demands for diversity should be considered to balance the relationship between accuracy and diversity.

Key words: recommended algorithm, residual, attention mechanism, collaborative filtering