Research On Hypernetwork Link Prediction Method Based On Multidimensional Feature Mining

Link prediction is an important issue in network science in recent years.The core is to use the system architecture built by the network to carry out the research of multi-dimensional behavior patterns and the mining of long-term evolution mechanisms,so as to accurately predict unknown or future links.In the era of big data,with the extension of space and the passage of time,more and more information is closely linked.A variety of complex systems show high-order and heterogeneous network characteristics.Ordinary complex networks are difficult to effectively extract higher-order network information and cannot provide a reasonable interpretation of the essential features of data analysis.As a scientific expression for characterizing such higher-order relationships,hypernetwork can more fully capture the topological properties and complex associations of relational data,deepening understanding and cognition of the complexity of the real world.The existing link prediction methods generally take the low-order network as the modeling object,ignoring the expression of the complete semantic information between multiple entities,leading to technical bottlenecks when the model performs feature mining and generalization migration for higher-order networks.Therefore,in view of the functional characteristics and structural commonalities of higher-order hypernetworks in different scenarios,this article classifies them and sequentially explores their internal mechanisms and evolution patterns.Then,based on the representation types of hypernetworks,different forms of data abstraction and feature space construction are carried out based on the representation types of hypernetworks,making full use of the connectivity principles of heterogeneous,attribute,and temporal hypernetworks to design prediction methods with strong universality,further improving the modeling ability and application range of chain path prediction.The main work of this paper is as follows:1.Aiming at the clustering characteristics of different hyperlinks in heterogeneous hypernetworks,a link prediction model based on hypergraph clustering parser is proposed.Firstly,the heterogeneous hypergraph is used to explicitly model the complex interaction between higher-order data,and a clustering parser is developed as a reasoning model to divide the complete hypergraph into multiple subgraphs with the same properties based on the similarity of hyperlinks.The hyperlink predictor acts as a generation model to complement the topology in the adjacency space of each subgraph.The behavioral characteristics of hyperlinks and higher-order structural patterns have been organically combined to achieve more accurate and comprehensive prediction.Nine sets of real hypernetwork validations show that the proposed model has significant improvement in prediction performance under AUC score,recall number,and recall rate indicators.2.Aiming at the data interaction between different modes in attribute networks,a hypernetwork link prediction model based on attribute and structure fusion is proposed.Firstly,a dual-channel encoder is designed to carry out collaborative learning of the structure and attribute characteristics of nodes.In particular,the convolutional principle of hypergraph neural networks is used to re encapsulate the representation of nodes and attributes,so that the structural information of nodes can be spread more through shared attributes.Finally,the decoder based on the hyperedge attention mechanism is combined to judge the hyperlinks that are more equivalent to the original network functions.The experiments of six groups of real attribute hypernetworks show that this method is superior to the comparison model in terms of the validity of hypernetwork representation and the accuracy of link prediction.At the same time,a large number of ablation experiments also prove the scientific nature of different module designs.3.Aiming at the spatiotemporal information under different vision in temporal network,a link prediction model based on temporal hypergraph dual attention network is proposed.The generation model takes the hypergraph convolution network that perceives spatial dependence and the gated recurrent network that focuses on time evolution as the core.The spatiotemporal attention mechanism has also been introduced separately to adaptively adjust the impact of different structural patterns on the temporal network.In addition,the game process of the discrimination model is added to the training of the generating model to continuously improve the presentation quality of future network snapshots.The experimental results in four real scenarios show that the proposed model has lower root mean square error and error rate than the baseline model.The ablation experiment also verifies the contribution of each component and attention mechanism to the improvement of model performance.