Research On Expert Finding Method In Social Q&A Community Based On Feature Analysis And Representation Learning

The popularity of the Internet has promoted the advent of information age,which enables people to access the massive information resources on the Internet.While bringing convenience to life,it also increases the difficulty of obtaining valuable information,resulting in the problem of information overload.In the information age,the diverse and personalized information needs and social needs of users have prompted the emergence of social question and answer site,which can provide the knowledge question and answer service as well as meeting the social needs of users.It is a new kind of knowledge question and answer service platform.Early social question and answer sites mainly use invitation or real-name registration to attract domain experts.With the development of communities,the influx of a large number of new users causes problems like long waiting time for new questions to be answered,or the lack of professional answers,these problems can lead to loss of users,which is not conducive to the healthy and sustainable development of social question and answer community.Thus,it is important to identify experts who are willing to answer questions in the community.Previous expert finding researches mainly focused on identifying users with high influence or authority in the community,which did not take into account users' willingness to share knowledge.The main purpose of this study is to identify experts who are more likely to answer questions in the community based on the analysis of knowledge sharing intention and behavior of users,and make up for the deficiency of previous expert finding researches.This research is conducted to solve the following questions:(1)How to analyze users' knowledge sharing intention and behavior under the guidance of knowledge sharing theories,and build a feature dataset for users based on the analysis;(2)How to extract the relationships between user and question through the deep mining of user's answer history;(3)How to use the feature dataset to identify experts who are willing to answer questions,and which features have important predictive value.As the first social question and answer community in China,Zhihu has accumulated a large number of users and massive question-and-answer data.This research takes Zhihu as the research object,a python crawler is used to collect data related to questions,answers and users under the topic of "psychology".Under the guidance of theories related to knowledge sharing,we analyzed users' knowledge sharing intention and behavior in Zhihu,we then extracted features based on the analysis,and constructed a feature dataset in three dimensions,namely user authority,user participation,and the relationship between user and question.In the process of feature extraction,in order to extract the relationship between users and questions from users' answer history,this study constructed a user-tag heterogeneous network,and used Node2 Vec to learn the embedding vector of nodes in the heterogeneous network,then cosine similarities were computed as the similarity features between users and question tags.Finally,machine learning classification models such as logistic regression,random forest and XGBoost were applied to the feature dataset to predict the probability of user providing high-quality answer and realize the effective identification of users who are willing to provide high-quality answers.The results of the empirical studies show that the XGBoost model performs best,achieving an accuracy rate of 0.70,a recall rate of0.46,an F1 value of 0.55,an AUC value of 0.99,and an AP value of 0.64.Comparing the expert list predicted by XGBoost with that of the classic Page Rank algorithm and HITS algorithm,it can be found that the algorithm proposed in this study takes into account both the users' professional level and willingness to share knowledge when identifying experts,and the predicted expert list is more consistent with the actual situation.By analyzing the feature importance of XGBoost,we found that the similarity features proposed in our study rank first,which make a great contribution to expert finding tasks.