| As an information filtering mechanism,the recommendation system can effectively alleviate the problem of "information overload" caused by the rapid growth of data volume.As the most commonly used technology in recommendation systems,collaborative filtering can effectively recommend products or services of interest to users from massive data.Due to the serious data sparsity problem in recommendation systems,that is,there are a large number of zero values or missing values in the interaction data between users and items,more and more researchers have begun to try to use auxiliary information to solve this problem.In recent years,the emergence of graph neural networks has promoted the development of graph embedding learning and provided a new research perspective for recommendation systems.Numerous studies have shown that graph neural networks have significant advantages in learning from graph data.Inspired by this,a research on the recommendation model based on graph embedding learning were carried out in this thesis.The main work and innovations are as follows:Aiming at the sparse problem of behavioral data in traditional recommendation systems,a bilinear diffusion graph recommendation model based on the fusion of user social relations were proposed in this thesis.This recommendation model designs a bilinear diffusion aggregator based on graph convolutional network.The aggregator consists of two core parts:The first is the diffusion aggregation part used to capture user-neighbor interaction information,which dynamically models the social influence of users and aggregates neighbor information from the local neighborhood of the social graph to the target.The second is the bilinear aggregation part used to capture neighbor-neighbor interaction information.The function of this part is to capture the potential interaction signal between neighbors and use it as user auxiliary information to improve user embedding.In order to verify the effectiveness of this model,this thesis designs an experiment and compares it with existing recommendation models.The experimental results show that the model has higher Hit Ratio and Normalized Discounted Cumulative Gain than existing recommendation models.Therefore,the bilinear diffusion graph recommendation model integrating user social relations can effectively alleviate the sparse problem of user behavior data,and greatly improve the recommendation accuracy.Aiming at the problems of poor model generalization ability and insufficient robustness in recommendation systems.From the perspective of adversarial learning,a directed adversarial graph convolution recommendation model based on neighbor differences were proposed.A concise and effective social regularization method were designed and introduce it into the recommendation model in this thesis.The model reconstructs the user’s adversarial instances in the discrete social graph domain,and injects perturbations into the training process of user embeddings to enhance the model’s social synergy signal,which reflects the similarity of user’s interest in the social network.The model achieves the entire user regularization process by directional interference with user feature embeddings.This thesis conducts comparative experiments with the existing recommendation models.The experiments show that the directed adversarial graph convolution recommendation model based on neighbor differences is less prone to overfitting problems,has stronger generalization ability,and has higher recommendation accuracy. |
