Link Prediction In Dynamic Heterogeneous Graph Based On Graph Neural Networks

With the development of information technology,a large amount of graph data is accumulated,a considerable part of which is dynamic heterogeneous graph data.A dynamic heterogeneous graph is composed of different types of events with temporal labels.Dynamic heterogeneous graphs are applied to several downstream tasks,such as link prediction,node classification.Among these tasks,link prediction for dynamic heterogeneous graphs is widely used in many important scenarios such as recommendation systems and social networks.In recent years,graph neural networks have made great progress in performing the link prediction on dynamic heterogeneous graphs.However,there are still some questions of concern in link prediction for dynamic heterogeneous graphs,including but not limited to(1)how to effectively exploit the decaying impact of events,(2)how to efficiently utilize the causality of event types.The aforementioned issues are investigated in the dissertation.And we summarize our contributions as follows:(1)Research on how to effectively exploit the decaying impact of events.This dissertation firstly proposes the event decaying mechanism on the dynamic heterogeneous graphs.Secondly,this dissertation proposes the Event-Decay-Based Heterogeneous Attention Network(DHAN)based on the event decaying mechanism.DHAN utilize the proposed Event-Decaying-Based Aggregation Module(EDBA).EBDA combines the event decaying mechanism and an extension of the graph self-attention mechanism.While computing the attention of nodes to their neighbours,EBDA assigns higher attention to events with less decay,making the aggregated information more time-sensitive.In addition,by utilizing a temporal forgetting mechanism to aggregate information from graph snapshot sequences,DHAN enables personalised event impact calculations and captures the temporality of dynamic heterogeneous graphs.(2)Research on how to efficiently utilize the causality of event types.Firstly,this dissertation explicitly defines causality of event types by Heterogeneous Causality Graph(HCG)to utilize causality from the perspective of graph structure.By masking the events of irrelevant types under the guidance of HCG during the prediction,the causality of event types is utilized.Secondly,this dissertation proposes the Event-Type-CausalityBased Continuous-Time Heterogeneous Attention Network(ECHN)to model dynamic heterogeneous graphs.To utilize the causality of event types from the perspective of modeling algorithm,ECHN aggregates features based on the strength of different causal relationships between event types in the prediction process.ECHN assigns greater attention for events with highly relevant types.The utilities of causality of event types reduces the noise from irrelevant events.This dissertation validates the effectiveness of DHAN and ECHN on several datasets,and the results show that above methods achieve better performance on link prediction experiments compared to baselines.