Heterogeneous Information Network Representation Learning And Its Application In Recommendation Algorithms Research

With the rapid development of computer science and artificial intelligence,research on graph neural networks has shown an exceptionally active trend.Graph neural networks can map real-world problems to nodes and edges in a graph network,and efficiently process graph-structured data by modeling relationships between nodes.The emergence of this technology greatly improves the efficiency of handling graph-structured tasks such as social networks and recommendation systems.As an important branch of graph neural networks,heterogeneous information networks have wide application prospects in practical problems due to their inclusion of different types of nodes and edges.This paper aims to explore the representation learning methods of heterogeneous information networks and their practical applications in recommendation systems.Specifically,we need to address two core issues:1)how to find an effective method to embed nodes in heterogeneous information networks as low-dimensional vectors while preserving heterogeneity information as much as possible;2)how to use these embedding vectors to construct a recommendation system model that generates more accurate recommendations.The main contents of this paper are as follows:(1)Representation learning method based on heterogeneous information networks.We propose a fusion encoding and adversarial attack meta-path aggregation graph neural network(FAMAGNN).The model consists of three module components.namely,node content transformation,intra-meta-path aggregation,and inter-meta-path aggregation,which aim to solve the problem of insufficient feature extraction in existing heterogeneous information network representation learning methods.At the same time,the model introduces a fused meta-path instance encoder to extract rich structural and semantic information in the heterogeneous information network.In addition,we introduce FGM adversarial training to perform adversarial attacks during model training to improve the robustness of the model.The outstanding performance in downstream tasks such as node classification and node clustering proves the effectiveness of this method.(2)Recommendation system based on heterogeneous information networks.We propose a recommendation model based on attention mechanism and heterogeneous information network meta-paths(AMMRec).The model enhances the representation ability of feature vectors of nodes such as users and projects by using an attention neural network to fuse direct and indirect information,thereby improving the recommendation performance.AMMRec designs an attention neural network to fuse the representation vectors of different meta-paths in a mutually reinforcing manner,improving the feature representation of users,meta-paths,and projects.Comparative experiments on two real datasets shows that AMMRec model is effective in improving the performance of recommender systems and has the potential to solve the cold start problem of recommender systems.