Design And Empirical Analysis For Multi-scenes Recommendation Models Based On Graph Attention Networks

Recommender systems have been extensively applied to many portal websites and internet applications due to high-efficiency user conversion rates,the ability to attract and retain users.With traditional recommendation algorithms gradually tend to be saturated and deep learning technologies have begun to develop comprehensive technological innovation in the field of recommendation.Under the background of the digital age,only using users’ historical behavior information for recommendations can not achieve the strategic goals of enterprises.The prior information is hard to collect in real recommendation scenarios,and most available data exhibit non-Euclidean manifold properties,while traditional deep learning techniques are not applicable.As an emerging technology in the artificial intelligence field,graph neural networks have effectively solved this problem by mining complex relationships in graph structures.According to the different prior information in real recommendation scenarios,this thesis designs two reasonable and efficient recommendation models based on graph attention network theories that can simulate the different importance of neighbors to the center node.The primary work includes the following aspects:For the recommendation scenarios containing multiple prior graph data,a recommendation model based on multi-graph attention fusion is proposed to provide users better recommendation experience.The dual-branch residual graph attention module is presented to extract neighbors’ similar features from user and item graphs effectively and easily.Multi-scale latent matrices are captured by applying non-linear transformations to reduce the cost of dimension selection.Furthermore,a hybrid fusion graph attention module is designed to obtain valuable collaborative information from the useritem interaction graph,and refine the latent features of users and items.Finally,the whole recommendation framework is optimized by a geometric factorized regularization loss.Extensive experiment results on both synthetic and real-world datasets illustrate that our model can achieve better recommendation performance with a certain level of interpretability than some existing approaches.For recommendation scenarios lacking some prior graph data,a deep dynamic graph attention recommendation framework based on the influence and preference relationship is developed to promote the long-term retention of users.Since this scenario lacks the user graph or item graph,graph structures are reconstructed from two aspects:influence information and preference information.When designing two dynamic graph attention modules,we break fixed constraints of initial graph structures by introducing reconstructed influence graphs and constructed preference graphs.In addition,techniques of dropping and keeping edges are used to adjust initial graph structures and a deep feature aggregation block and adaptive feature fusion operation are introduced to obtain high-order information.The design of the adaptive fine feature extraction module enables the model to capture finer latent features.Experiment results on some benchmark datasets indicate the effectiveness and superiority of our model over some existing recommendation models.