Research And Implementation Of GNN Algorithm Based On Disentangled Representation

In real-life scenarios,there are lots of graph data formed by entities(object nodes)and their interconnected relationships(edges).Researchers out classical machine learning tasks(node classification,link prediction,clustering,etc.)from informative graph data,and then solve many complex but practical problems(object detection,recommendation systems,social analysis,etc.),thus graph analysis has become a research hotspot in the field of data mining and machine learning.In recent years,Graph Neural Networks(GNNs),as a powerful deep representation learning method for modeling graph data,has rapidly aroused considerable research interest in various fields.GNNs capture dependency of graphs through message passing module between graph nodes,which can well integrate structure and attribute information in data,and those rich variants constantly achieve performance breakthroughs in different tasks.More generally,many complex systems are often composed of different types of components and diverse relationships,and Heterogeneous Graphs(HGs)can describe them more completely and naturally without resulting in lots of information loss.However,the traditional GNNs are insufficient to distinguish and express the heterogeneity of objects and their relationships,thus the corresponding Heterogeneous Graph Neural Networks(HGNNs)emerge,which promote the research of graph data mining into a new era of development.Even though GNNs(including homogeneous and heterogeneous graphs)have achieved convincing performance,they have rarely further explored the underlying multi-facet factors behind the intricate interactions of graph data.It is difficult to obtain definite attributes or labels for a large number of interactions in graph data,but there are still many implicit factors behind these interactions.If multi-facet factors are failed to identified and disentangled in the process of information propagation on GNNs,the learning highly entangled and single representations will greatly reduce the robustness and interpretability of the model.To solve this problem,this paper proposes GNN algorithms based on disentangled representation,and analyzes how to perform disentangled representation learning in GNNs in different practical scenarios,including two research works:firstly this paper focus on the news recommendation system under the graph disentangled neural network,proposes a news recommendation algorithm named GNUD(Graph Neural News Recommendation with Unsupervised Preference Disentanglement)that can perform unsupervised disentanglement of user preferences;secondly,this paper studies the disentangled learning method in scenarios with more general text-related heterogeneous graph data,and proposes a topic-aware HGNN for link prediction.In news recommendation,the traditional methods usually utilize the user’s historical interactions and news content for recommendation prediction,ignoring the high-order structural relationships of the interactions.In addition,existing works failed to identify and disentangle users’latent but diverse preferences which cause theirs clicks on different news.Therefore,in this work we proposes a novel Graph Neural News Recommendation with Unsupervised Preference Disentanglement,named GNUD,which can identify and disentangle neighbor information while using higher-order structural relationships to propagate it,and discriminately aggregates neighbor attributes.Experiment results show that GNUD effectively improves the performance compared with other state-of-the-art methods on the real news recommendation datasets.For more common scenarios with heterogeneous graph,although existing HGNNs are able to capture rich semantic information and reveal different aspects of nodes to a certain extent,they still stay at a coarse level that simply exploits structural characteristics.In fact,a large number of nodes in HGs carry rich unstructured text,containing latent but more fine-grained semantics arising from multi-facet topic-aware factors.The factors can fundamentally manifest why nodes of different types would connect and form a specific heterogeneous structure.However,little effort has been devoted to factorizing them.Therefore,this work proposes a Topic-aware Heterogeneous Graph Neural Network,named THGNN.Compared with existing methods,this model can hierarchically mine topic-aware semantics for learning multi-facet disentangled representations for link prediction in HGs.Experimental results on real HGs demonstrate that THGNN not only outperforms existing state-of-the-art methods in the link prediction task,but also shows the potential interpretability of learnt multi-facet topic-aware representations.