Multi-feature based Question–Answerer Model Matching for predicting response time in CQA

• Published in: Knowledge-based systems Vol. 182; p. 104794
• Main Authors: YueLiu, Tang, Aihua, FeiCai, Ren, Pengfei, Sun, Zhibin
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 15.10.2019; Elsevier Science Ltd
• Subjects: Algorithms; Community Question Answering; Dirichlet problem; Feature extraction; Model matching; Predictions; Questions; Response time; Search engines; Stack overflow; The prediction of question response time; Topic-sensitive model; Community Question Answering; The prediction of question response time; Topic-sensitive model; Stack overflow
• DOI: 10.1016/j.knosys.2019.06.002
• ISSN: 0950-7051

Users of Community Question Answering (CQA) could not manage their time conveniently because their questions are often not answered quickly enough. To address this problem, we try to provide a function for CQA sites to inform users when their questions will be answered. In this paper, we propose a Question–Answerer Model Matching based answerer’s response time prediction named (QAM2), which consists of two parts: the construction of the Multi-feature based Question–Answerer Model (MQAM, including the answerer model and the question model) and the prediction of question response time based on MQAM Matching Strategy (QAMMS). Firstly, the MQAM is built according to some extracted deep features (e.g., answerer’s interest, professional level, activity, question category and difficulty), which are neglected in most existing methods on the prediction of question response time. Herein, the Label Cluster Latent Dirichlet Allocation (LC-LDA) model was proposed to overcome the compulsive allocation behaviors caused by traditional topic models (e.g. LDA), which treats the words that are irrelevant or weakly related to the subject as the topic of short texts when extracting the feature of answerer’s interest and question category. Meanwhile, an improved PageRank algorithm-topic sensitive weighted PageRank (TSWPR) is used to eliminate the impact of “indiscriminate” users who have answered many questions with low quality of answers. Secondly, we use the model matching strategy based on multiple classifier for matching MQAM and calculating the question response time of each answerer. Experiments conducted on two real data sets of Stack Overflow show that the proposed method can improve significantly the accuracy of question response time prediction in CQA.