Research On Expert Recommendation And Text Classification Based On Heterogeneous Network Embedding

With the rapid development of artificial intelligence,deep learning began to shine in many fields,such as computer vision,natural language processing,recommendation system.However,there are many scenes with sparse data,where deep learning is difficult to achieve good results because there are not enough training samples.Expert recommendation problem and super multi categories text classification problem are two typical data sparsity scenarios.The former requires high user interest and expertise level for user's questions and answers,which makes the interaction between users and questions relatively sparse.For the latter,there are too many categories,so the data among categories is unbalanced,and the data of most categories is relatively sparse.The key to solve the problem of data sparsity is how to enhance the transmission and sharing of information by introducing more information and association,which is the advantage of network structure.Therefore,this paper uses heterogeneous network to model these two problems,and uses network embedding method to extract and utilize the rich information on heterogeneous network,so as to promote the recommendation and classification performance.The main research contents and innovative achievements of this paper are as follows:1.Aiming at the problem that it is difficult to accurately describe the user's interest and expertise in expert recommendation task due to data sparsity,this paper proposes an expert recommendation method based on heterogeneous network embedding.First of all,in order to alleviate the problem of data sparsity,a heterogeneous network composed of users and problems is constructed.The method based on meta-path random walk is used to learn users' long-term interests,which is combined with the short-term interests of users using long-term and short-term memory network.Then,in order to model user expertise explicitly,a feedback aggregation network is proposed,which uses the user's historical answers and feedback information to get the user expertise.By using the long and short-term memory network with attention mechanism,this network can also capture the dynamic changes of user expertise and make the representation of user expertise more accurate.On the stackexchange public dataset,the proposed algorithm improves each index by nearly 4%.2.Aiming at the problem that with the increase of the number of classification categories,the model can not learn the effective representation for the categories with sparse data,which leads to the low classification accuracy,this paper proposes a text classification method based on heterogeneous network embedding.This method uses heterogeneous network to model text data,constructing words,documents and categories as text heterogeneous network,and learns the representation of categories explicitly.In order to capture node attributes and network structure information at the same time,and make good use of tags,this paper designs heterogeneous graph neural network on the constructed heterogeneous network to learn node representation.Considering the heterogeneity of nodes,different aggregation functions are used for different types of nodes,so that information can be effectively transferred in the network.Taking into account the hierarchical association between categories,a multi-level classification loss is designed to model the association between categories into heterogeneous networks,introducing more abundant information to reduce the difficulty of classification.This method improves the accuracy of nearly 3%on the telecom customer service dataset,verifying the effectiveness of heterogeneous network modeling.