Research And Application Of Attribute Network Embedding Based On Sparse Autoencoder

With the explosive growth of Internet data,the multi-source,heterogeneity,randomness and fuzziness of massive data sources have brought new challenges to data analysis.How to mine useful information from massive data and provide basis for enterprise decision-making has become increasingly important,and has gradually become a new research topic.There are many methods to mine effective information from massive data.At present,one of the mainstream processing methods is to express data as attribute network,then extract network features by means of network embedding,and finally use the extracted network features to complete specific tasks,such as classification,clustering,link prediction,recommendation and so on.Network embedding is to encode nodes in the network into low-dimensional and dense vectors,so as to avoid the differences,high-dimensional and heterogeneity of large data sources and achieve the purpose of extracting intrinsic characteristics of the network.The research shows that extracting the intrinsic features of the network effectively can not only accelerate the training speed of the model,but also improve the accuracy of the followup tasks.This thesis studies how to extract network intrinsic features from the network,generate low-dimensional,dense vectors for network nodes,and verify the performance of network embedding on real networks.Aiming at the problem that the existing network embedding methods can not effectively mine the intrinsic characteristics of the network,this thesis focuses on the extraction of network embedding.The main research contents are as follows:(1)Weighted fusion of network topological features and semantic attribute information.The network topology part integrates three parts: the adjacency matrix of the network,the second-order neighbor extracted from the network topology and the common neighbor ratio information;the semantic attribute information is the modularity matrix of the semantic attribute calculated from the semantic attribute matrix.(2)Network embedding extraction.The semi-supervised sparse Auto-Encoding model is used to train the fused vectors,and the network embedding of attribute network is obtained.In the process of model training,the semi-supervised constraint and sparse loss constraint are added to the objective function to guide the model to extract the network feature process to get higher quality network embedding.(3)The attribute network embedding algorithm based on sparse Auto-Encoding applied to the field of citation recommendation.Using the citation relationship of HowNet documents,the network topology of HowNet attribute network is constructed;the semantic attribute matrix of HowNet attribute network is constructed according to the title,author,abstract,keyword,classification number and published periodical information of documents;the vector fused with network topology and semantic attribute matrix is fed into semi-supervised sparse Auto-Encoding odel to obtain network embedding;and the network embedding is used to aggregate.Class enables similar documents to be displayed adjacent to each other,so as to improve the efficiency of searching documents and achieve the purpose of accurate recommendation.