Research On Expert Recommendation Algorithm Of Online Question And Answer Community Based On Deep Learning

Online question and answer(Q&A)community is a web technology application with Q&A interaction as the core function.It can not only realize the sharing of community knowledge and experience but also provide users with timely and personalized knowledge recommendation services.With the continuous promotion of the Q&A community,the amount of data that the platform needs to process becomes huge and complex.In the face of massive data,it is difficult for respondents to identify the questions of interest,and questioners cannot obtain quality answers in a timely manner.Over time,a poor Q&A experience will lead to a decline in user engagement.In order to improve the efficiency of question answering and optimize the accuracy of question-user invitation,it is necessary to recommend questions to expert users who may answer them.Therefore,expert recommendation has become a big challenge for online Q&A communities.In terms of expert recommendation,the majority of existing works ignore the effective extraction of users’ dynamic interest representation,and the embedded representation of cold problems is poorly learned,resulting in insufficient accuracy in actual recommendation scenarios.This paper investigates the aforementioned issues and proposes an expert recommendation algorithm to address them.The main work and achievements are as follows:(1)Aiming at the problem of insufficient accuracy of expert recommendation in Q&A community caused by insufficient extraction of user’s dynamic interest representation,this paper proposes an expert recommendation algorithm for the Q&A community based on multi-head self-attention.Firstly,a question encoder was built using convolutional neural network and attention mechanism to deeply encode the question and extract the question representation containing deep semantic information.Secondly,the user’s historical answer sequence is regarded as a time series,and the dynamic interest representation contained in the multi-head self-attention mechanism learning sequence is used to obtain the user’s comprehensive interest representation combined with the user’s static interest representation.Finally,the target problem representation and user comprehensive interest representation are spliced and combined,and then combined with the user ability representation,the full connection layer is input for similarity calculation,and the recommendation result is generated.The algorithm considered the user’s dynamic and static interest representation and dynamically captured the user’s short-term interest changes according to the user’s historical sequence behavior,which made the recommendation results more realtime.(2)Aiming at the problem of insufficient expert recommendation accuracy in the Q&A community caused by poor representational learning of cold question embedding,this paper proposes an expert recommendation algorithm for the Q&A community based on cold question embedding learning optimization.Firstly,T-PageRank is used to obtain the global feature representation of the cold problem,and the time feedback factor is used to compensate the cold problem and keep it from sinking.Secondly,a reverse neighbor index table was built to quickly retrieve hot questions with high attribute similarity to cold questions,and an attribute neighbor graph was generated.Then,combined with the attribute feature of the cold problem itself,the graph attention mechanism is used to adaptively obtain useful information from the attribute neighbor graph to enrich the local feature representation of the cold problem.Finally,taking into account the global and local feature representations of the cold question,MAML is used to generate an ideal initial embedding representation for the cold question,which is then input into the depth prediction model with the user’s initial embedding representation for similarity calculation,and recommendation results are generated.The proposed algorithm effectively learns the cold problem’s high-level attribute feature representation and generates an ideal initial embedding for it,thereby alleviating the cold start problem and improving recommendation performance.In this paper,experiments are performed using the real Zhihu Q&A dataset.The results show that,when compared to other advanced algorithms,the proposed algorithm has certain advantages in dealing with existing problems recommended by experts in the Q&A community,and the recommendation performance is significantly improved.