Research On Node Classification Algorithm Based On Hypergraph Representation Learning

As a widely available data structure,graphs reflect the point-to-point connection relationships of various types of networks in the real world.However,pairwise connected graph structure relationships cannot effectively represent the higher-order complex non-pairwise relationships in the real world,so hypergraphs as generalized representations of graphs have received extensive attention.However,hypergraph-based neural network models usually focus on learning the representation of hypergraphs in first-order neighborhoods,while ignoring node interactions in higher-order neighborhoods.In addition,missing information is a common situation in the real world,such as inaccessible personal sensitive information in social networks and missing data corruption in sensor networks.Current hypergraph representation learning usually assumes that nodes have complete attribute information,i.e.,complete features,and does not consider node classification in the case of missing features.Therefore,this thesis addresses the above-mentioned problems and researches the higher-order neighborhood interaction problem in hypergraphs and the node feature missing problem in hypergraphs from the node classification task of hypergraph representation learning,respectively.The main work and research results of this thesis are as follows:（1）We propose the 2-order and 3-order neighborhood approximation spectral hypergraph convolution operators and Hypergraph Convolutional Network with Hybrid High-order Neighbors（Hybrid HGCN）model for the node classification task with complete feature cases.A hypergraph convolution module based on 2-order and 3-order neighborhood approximation spectral hypergraph convolution operators is derived using truncated Chebyshev polynomials with K-th order restrictions on the hypergraph neighborhood order.On this basis,we design a hypergraph convolution network with hybrid higher-order neighborhoods in the form of multichannel to fuse the hypergraph neighborhood information of different orders in the nodes effectively.Experiments show that the neighborhood approximation spectrum hypergraph convolution operator proposed in this thesis achieves node and long-range hyperedge information fusion;the hypergraph convolution network model with multi-channel hybrid strategy effectively uses high-order and low-order neighborhoods to achieve node classification.（2）We propose the Hypergraph Mixture Density Network（HMDN）model combining the Bayesian deep generation model for the node classification task of incomplete feature cases.The feature missing scenario is abstracted as a feature uncertainty problem,and the missing node features are represented using a mixture density function.In order to flexibly embed higher-order information in the mixture density,this thesis further introduces variational inference in the hypergraph mixture density,gives the log-likelihood function and Evidence Lower Bound（ELBO）of the hypergraph mixture distribution,and designs an inference and generation model based on the message propagation strategy of the hypergraph network.Experiments show that HMDN can effectively represent incomplete features with uncertainty and optimize the latent embedding of nodes;the inference-generation model based on a variational inference strategy can embed higher-order information in the mixture density and fuse the incomplete features with hypergraph topological information effectively.（3）We propose the Joint Learning of Hypergraph Mixture Density Network and Label Propagation（HMDN-LP）model for the node classification task in the case of a high feature missing rate.The hypergraph label propagation and embedding strategy on the HMDN model further enhance the multi-level representation of incomplete features by introducing label propagation.Meanwhile,two different label embedding fusion strategies are proposed,and mixture distribution log-likelihood function and ELBO are constructed under the label information condition.ablation experiments show that the label propagation strategy has a positive effect on the representation enhancement of incomplete features,and further improves the node classification accuracy in most feature missing cases.