Research On Link Mining Technology Based On Graph Representation Learning

Many scenarios in real life can be abstracted as graph structure which is composed of entities and relationships,such as traffic network,molecular structure,social network,etc.Graph mining can help relevant personnel analyze these complex structures and assist decision-making,which attracts scholars from many fields to study it.Especially with the development of graph representation learning and graph neural network,graph mining has become a research hotspot in recent years.Graph mining can be divided into node-level mining,link-level mining and subgraph or whole graph level mining.Link-level mining on graph is intended to discover the relationship or interaction between nodes,such as classical link prediction,but link mining has many other directions.Ordinary unweighted graphs cannot describe the strength of relationships between entities,and many real-world networks are vulnerable to abnormal links.Therefore,in this thesis,link weight prediction and anomalous link detection are selected from lots of link mining tasks,and the corresponding effective methods based on graph representation learning are designed.The main work and innovation are as follows:A self-attention enhanced auto-encoder for link weight prediction with graph compression on weighted graphs is designed which is called SEA.The attention mechanism,which has excellent performance in many tasks,is used to aggregate the neighbors of a node to obtain the node embedding vectors.Then a learnable node pair interaction function is as the decoder to reconstruct the weighted adjacency matrix,the results of it are used to predict the link weights.In order to alleviate the problems of video memory overflow and excessive resource occupation on large-scale graphs,a weighted graph compression algorithm for link weight prediction is proposed.The graph can be compressed before the prediction task.The experimental results on 7 real-world weighted network datasets demonstrate the effectiveness of the SEA model,and the graph compression algorithm can reduce the size of the graph while the damage to the local structure has little impact on the final prediction results.A robust graph structure embedding model,RGSE is designed to detect anomalous links on static graphs.Existing graph representation learning and graph neural networks assume that edges in the graphs are trustworthy.However,the addition of abnormal edges will lead to the deviation of node embedding vectors.Especially the graph neural networks based on message passing are easy to be affected by abnormal edges,which will have a negative impact on the results of the model.On the one hand,the RGSE model uses a node-level contrast loss in the auto-encoder to constrain the embedding vectors and obtain more robust node embeddings,then the representation of links can be generated.On the other hand,since the second-order structure between nodes is relatively stable and not susceptible to abnormal edges.Therefore,feature extraction of node pairs is performed based on the second-order structure.Then,the features of the two parts are fused and the self-supervised learning is carried out according to the adjacency relationship.Finally,anomaly detection is performed according to the link existence probability.3 different anomalous links are injected into 5 data sets,and the experimental results show that RGSE has good detection effect.Last but not least,the robustness and effect of node-level contrast loss constraint and second-order structure features of node pairs are analyzed in detail.