Research And Application Of Link Prediction In Heterogeneous Information Network Based On Transfer Learning

Link prediction is a popular research direction in the field of data mining,which aims to predict the possibility of potential links between network nodes through known network structure information.Link prediction can effectively support applications such as script event prediction,network structure knowledge mining,and network evolution pattern discovery.Current research on link prediction mainly focuses on homogeneous information networks,which consist of only one type of node and link.However,little attention has been given to heterogeneous information network scenarios that are ubiquitous in the real world.In heterogeneous information networks,more abundant types of nodes and links bring new challenges to the existing link prediction methods and applications for homogeneous information networks.The challenge is mainly reflected in how to mine transferable knowledge from heterogeneous information networks to reveal the evolution rules of different link types.Therefore,this thesis researches link prediction technology and its application for heterogeneous information networks.A transfer learning-based link prediction model for heterogeneous information networks is proposed,which employs domain adversarial networks to enhance the transferability of task-related knowledge.Based on this model,further application research on link prediction is carried out,and script event prediction is selected as a typical application demonstration.The main research work of this thesis is summarized as follows:First,this thesis proposes a new transfer learning-based link prediction model for heterogeneous information networks,called the Transferable Domain Adversarial Network(TDAN).Existing link prediction models can learn type-specific link knowledge through feature extraction.However,since different link types frequently exhibit distinct evolution patterns,existing methods usually suffer from poor generalization,especially when predicting potential links on long-tail datasets or new link types.In order to alleviate this problem,on the one hand,TDAN employs adversarial training to maximize the domain similarity of features across different link types and aims to automatically transfer the learned task-related knowledge to the target link type.On the other hand,to alleviate the possible noise problem in the process of capturing type-shared knowledge,TDAN adopts an autoencoder to enhance the network structure knowledge of link features.The experimental results on multiple public datasets of Facebook and You Tube demonstrate that TDAN significantly outperforms all the baseline models involved in the comparison.The ablation experiments and related analysis also fully verify the feasibility and effectiveness of the proposed model.Second,this thesis proposes a new model for the application of heterogeneous information network link prediction on scripted event prediction.Script event prediction aims to select the next possible event in the list of candidate events based on a given number of historical events.Due to the diverse link types present in candidate events,this application can be intuitively modeled as a typical link prediction problem in heterogeneous information networks.Specifically,the given continuous heterogeneous historical event nodes naturally contain the logical knowledge of the evolution of the event chain,which has an important guiding significance for the prediction of subsequent events under the same event chain.Existing script event prediction models rely on external knowledge to enhance event features,but external knowledge is often difficult to obtain and lacks certain generalization capabilities for different event chains.To alleviate this problem,this thesis further proposes an internal knowledge transfer network model(IKTN)based on the TDAN model.On the one hand,IKTN captures the implicit knowledge between historical events through the multi-head self-attention mechanism and graph convolutional network,and transfers it to the features of the given event chain and correct candidate event;on the other hand,IKTN adopts the sequence model to enhance the contextual semantic information of event representations while alleviating the noise introduced during the process of transferring internal knowledge.The experimental results on the NYT public dataset show that the performance of the IKTN model is significantly better than the baseline model involved in the comparison,and the ablation experiment further proves the feasibility and effectiveness of the proposed model in the application of script event prediction.