Research On Key Technologies Of Heterogeneous Graph Neural Network

The real-world data usually come together with the graph structure,such as social graphs,e-commerce graphs,academic graphs,and the world wide web.Graph-structured data is able to flexibly model the diverse interactions in complex information system,which has been widely used in many areas.Thus,graph mining has aroused considerable research interest and has become one of the most important directions in data mining.As a matter of fact,the real-world graph usually comes with multi-types of nodes and edges,also widely known as the heterogeneous graph.How to explore and exploit the potential value in heterogeneous graphs is a valuable research direction.Recently,graph representation learning(GRL)especially graph neural net-work has roused considerable research interest in graph mining.GRL aims to project the nodes,edges,and graphs into a low-dimensional space and preserves the graph structure and property.Heterogeneous graph neural network(Het-eGNN),as a powerful deep representation learning method for graph-structured data,is able to capture complex structure and rich semantics and improves the performance of downstream tasks.Taking recommendation system as an ex-ample,by considering the diverse interactions between users and items,Het-eGNNs capture the multiple potential preferences of users and then improve the accuracy and diversity of recommendation results.Considering the heterogeneity of graph,when designing architecture for HeteGNN,we have the following challenges:(1)Howto full consider the mul-tiple types of nodes/edges and rich semantics to design graph neural network architecture.(2)How to alleviate the degradation phenomenon in deep Het-eGNN.(3)How to improve the representational power of HeteGNN.To over-come the above challenges,this thesis mainly investigates the key technologies of HeteGNN,including heterogeneity modeling,degradation phenomenon,and representational power.In summary,the main contributions and innovations of this thesis are shown as follows:First,we study the problem of heterogeneity modeling and propose a hi-erarchical attention based heterogeneous graph attention network(HAN).It mainly consists of node-level attention and semantic-level attention.Specif-ically,the node-level attention aims to learn the importance between a node and its meta-path based neighbors,while the semantic-level attention is able to learn the importance of different meta-paths.Extensive experiments show the effectiveness and explainability of HAN.Second,we discover the degradation phenomenon(called semantic confu-sion)in deep HeteGNN and design a deep heterogeneous propagation network.Via theoretically analyzing the degradation phenomenon,we propose a het-erogeneous graph propagation network including semantic propagation mech-anism and semantic fusion mechanism,which is able to absorb the character of each node and fuse rich semantics.Third,we point out the limited representational power of HeteGNN and improve it from the perspective of relative distance modeling.Due to limited representational power,HeteGNN fails to distinguish and embed isomorphic graphs.Specifically,we propose heterogeneous shortest path distance to cap-ture the relative distances between nodes and then injects them into the neighbor aggregating process,significantly improving the representational power of Het-eGNN.Experimental results including transductive and inductive link predic-tion demonstrate the representational power and generalization ability of het-erogeneous distance encoding.Fourth,we apply HeteGNN to share recommendation and combination recommendation.Share recommendation aims to recommend a most likely friend to a user who would like to share a specific item.To solve it,we pro-pose a HeteGNN based share recommendation HGSRec with delicately de-signs:tripartite HeteGNN,dual co-attention mechanism,and triplet representa-tion.Combination recommendation aims to select a set of K items based on all user preferences and maximize the total revenue for the promotion scenario.On the basis of user/group preference estimated via HeteGNN,we propose the comb-K recommendation model,a constrained combinatorial optimization model with delicately designed constraints.Experimental results on share sce-nario and promotion scenario how the superiorities fo the proposed HeteGNN based recommendation models.