Personalized Recommendation Algorithm With Attention Mechanism

With the rapid development of Internet technology,"information overload" has become a problem that traditional search technology can't do.In order to solve this problem,the recommendation system came into being.Collaborative filtering algorithm is used widely and it is the successful technology in the recommendation system.However,the collaborative filtering algorithm is faced with the problem of serious data sparse and cold start.In recent years,with the enrichment of social information acquisition methods,the recommendation algorithm integrating user and project socialization has been widely used,which alleviates the problem of too sparse traditional user project interaction matrix to a certain extent.However,the existing methods fail to consider the differences of users' social status and trust objects in different fields,and the calculation of user similarity can not adapt to different projects,resulting in the decline of algorithms' accuracy.On the other hand,with the extensive application of deep learning,various neural network models have also been applied to the recommendation system,benefiting from its' strong ability of deep feature learning,the recommendation system has made great progress.However,it is not advisable to increase the number of neural network layers and the complexity of the model to obtain higher recommendation accuracy.In recent years,some researchers have introduced the attention mechanism which is very popular in NLP field into recommendation algorithm and achieved some results.However,the existing study model is too single,which only integrates the attention of users and other unilateral information,and fails to fully mine the information of social context and project level.Therefore,using attention mechanism to the model of multi domain information in recommendation problem has become the focus of this study.The main contents of this paper are as follows:(1)The existing methods can't consider the differences of users' social status and their trust objects in disparate fields,and similarity between users can not change automatically in the face of different projects.To address the issues,a collaborative filtering algorithm combining project information and trust mechanism is proposed.Items are divided according to their fields.Trust network of the specific field is built by considering the global trust and local trust of users in different fields.This paper proposes a new way to construct social information.Then,the similarity between items is integrated into Pearson correlation coefficient to calculate the degree of user's preference when facing different items.Experimental results using the Epinions dataset show that the performance of the proposed algorithm is greatly improved,compared to the classical collaborative filtering algorithm and the algorithm incorporating single information.(2)To solve the problem that the existing attention mechanism is too single in the process of recommendation,and the traditional social recommendation can not fully mine the nonlinear information in the social matrix,this paper proposes a recommendation method integrating multi-level attention mechanism.In this method,the multi-dimensional information is represented by Denoising Autoencoder and attention mechanism is integrated into users,project attributes,project content and social information at the same time.At the same time,the method also considers the integration of user project basic implicit vector,and uses multi-layer perceptron at the top of the model to make score prediction.Experimental results using the Flixter and last.fm dataset show that the performance of the proposed algorithm is greatly improved,at the same time,this paper talk about the difference between the different parameters in the experience.