Link Prediction Considering Text Sentiment Information

With the development of Internet technology,social platforms represented by Weibo,Twitter,Facebook,etc.and life consumption platforms represented by Yelp and others have gradually formed a complex network containing social attributes.These networks contain various and volume data.Mining the information contained in these data is of great significance for improving platform functionality and improving user experience.For example,Weibo,QQ,etc.can recommend friends to users by mining potential links existing between network nodes.Link prediction is based on the known network structure and node properties to predict the possibility of a connection between a pair of nodes in the network that have not yet been connected.The typical method is mainly to predict through network structure information.However,when the network is sparse or the network structure is relatively low,the prediction based on the network structure features cannot provide sufficient data information,which leads to a significant degradation of the performance of the prediction algorithm.This paper proposes a link prediction method that combines textual sentiment information to compensate for the lack of network structure information.Firstly,the text emotion feature extraction model is constructed to extract the emotional features contained in the text data of the network node.Secondly,the extracted features are merged with the network structure features and input into the classifier for link prediction,thus the problem of insufficient information in sparse networks is solved to some extent.The main work and innovations of this paper are as follows:(1)Text emotion feature extraction.First,the word features are designed according to the emotional polarity and part of speech of the words,and then each word in the text is mapped into an emotional word vector according to the word features.Secondly,a convolutional neural network is introduced for text emotion feature extraction,and text represented by the emotion word vector is taken as input.Finally,the attention mechanism is added before the convolution operation,which emphasizes the connection between the texts of each pair of nodes,then adding attention mechanisms after the convolution operation to distinguish the contribution of different words to the text representation.(2)Network structure feature extraction.The experiment compares the feature extraction ability of four network embedding algorithms: Deepwalk,Node2 vec,LINE and SDNE.Finally,the Node2 vec algorithm is selected as the network structure feature extraction method.(3)The combination of network structure features and text emotional features.First,the obtained network structure feature vector and text sentiment feature vector are fused by vector concatenation to obtain a final node feature vector.Then,the link feature vector is calculated according to the node feature vector,and the link existing in the network is taken as a positive sample,and the non-existent link is taken as a negative sample.(4)Link prediction using fusion features.In this paper,the link prediction problem is regarded as the binary classification problem of the link,and the merged feature is input into the SVM classification model for link prediction.The experimental results show that compared with the link prediction method based on network structure information,the proposed method not only greatly improves the prediction accuracy but also solves the problem of predictive performance degradation caused by insufficient network structure information in sparse networks.