Multi-label Learning Algorithms Based On Local Pairwise Label Correlations And Its Application In Zhihu

With the rapid development of Internet economy,knowledge sharing pattern has gradually developed into the transmission mode of information resources.Knowledgebased Q&A community is a knowledge-based social platform,which aims at the matching and sharing of information resources between suppliers and consumers.Questioners issue questions and give the corresponding labels.Answers retrieve questions by labels and answer them.Successful match depends on the platform's content distribution function and information retrieval system.Accurate tagging can improve users' experience and reduce operation costs.At present,the topic of knowledge is labeled by the user based on their thinking.This labeling method will cause the problem that the label is not associated with the problem or not accurate enough,which will reduce the efficiency of content distribution and user experience of the platform,and result in a certain amount of manual verification costs.Therefore,it's necessary to design an efficient topic tag automatic annotation algorithm in order to both improve the efficiency of content distribution and reduce human costs.The automatic tagging of question on knowledge-based social platforms is essentially a multi-label text categorization task,among which each problem may correspond to multiple tags.The problem of multi-label text categorization mainly includes two difficulties.One is that there are noise and errors in the process of text feature extraction,which reduces the performance of classification model.The other is that there are many complex semantic relationships among different labels in the multilabel tasks,which requires an efficient multi-label classification algorithm.Deep learning has significant advantages in automatic feature extraction and is widely used in the field of natural language processing.RAkEL is a high-order multi-label classification algorithm,which can model high-order label correlations,but it ignores label combination correlation when constructing the multi-label classifier.Therefore,this paper proposes an improved k-labelset algorithm based on local pairwise label correlations LPCkEL,combined with text feature extraction technology of deep automatic encoder,in order to realize automatic tagging of question.LPCkEL captures the dependences or exclusive semantic relationship among different text tags by means of k-nearest neighbor method and matrix similarity,and presents two optimizing strategies including the improved k-labelset generation method and rectification mechanism in order to predict more accurate and more complete labelsets.Finally,in this paper,a series of multi-label classification experiments are designed based on Knowledgeable Social Platform.The experimental results show that LPCkEL can achieve significantly better performance than some other the-art-of-the-state multilabel classification algorithms,and both strategies can improve the classification performance of RAkEL.