Research On Recommendation Algorithms In Online Community Question Answering

The community question answering community has gradually developed into a platform for people to share and acquire knowledge and information.Every day,a large number of new questions are asked by users,waiting for the answers and discussions of the other users.However,as the development of the community has accumulated a huge amount of problem,answer and user data.CQA websites has begun to face the problem of "information overload".On the one hand,users are difficult to quickly find problems they are interested in.On the other hand,a lot of new problems have been hidden in huge amounts of data,people can't get high quality answers in time.Moreover,many new problems lack of topic labels which can accurately describe the problem,as a result,they are hard to be retrieved by the rest of the users.This thesis aims to solve the problem of data challenges faced by CQA websites,and proposes the algorithm model to solve tag recommendation and expert user recommendation.The research work of this thesis include two parts.The first part puts forward the label recommendation algorithm based on deep learning.According to the multi-labels of problem,firstly define tag recommendation as a multi-label text classification,and then combine bi-directional long short-term memory network(Bi-LSTM)and convolutional neural network(CNN)to extract text semantic characteristics of the information,and train model in a supervised way from the training data.In order to improve the performance of the model,this thesis introduced the word attention and sentence attention mechanism based on the traditional attention mechanism in the bidirectional long short-term memory network.The second work of this thesis is propose an algorithm of expert user recommendation for new questions in CQA websites,this article defines expert users recommend as a pairwise learning problems,that means for each problem,according to the quality of the answers,create the partial order for every two answers to train model.In order to alleviate the sparse of user behavior and to enhance the quality of user-problem matching,this thesis also construct a heterogeneous graph based on user behaviors of answering the question and the user community social relationship,and then find more user-problem matching relationship through random walk in heterogeneous graph.This part apply Bi-LSTM learn represenation of problem,at same time learn a user embedding matrix to learn to rank.In order to enhance the expression learning ability of neural network,a multi-topic attention mechanism is proposed for the multi-topic attribute of CQA websites.Experimental results on zhihu datasets demonstate proposed algorithms outperform previous work of label recommendation and expert user recommendation,F1 – Score of the label recommendation outperform the traditional based on the content method outperform by 30% and 10% than single deep learning model.The expert user recommendation algorithm improved by about 10% on NDCG and MRR,and increased by more than 3% on the F1-score index than previous works.