Research On Multi-source User Behavior Analysis Based On Deep Learning

In recent years,with the rapid development of computer information technology and applications,such as mobile Internet and Internet of things,millions of users in various Internet media are producing a variety of interactive behavior data all the time.These data are not only very huge,but also contain various forms of data.This thesis focuses on the user generated interactive behavior data to analyze the user 's behavior comprehensively.The typical key issues are studied in this thesis,including the user's click stream behaviors,text classification,image recognition and video action recognition.Specifically,our main research contents and work are summarized as follows:(1)For user 's click stream,a two-level clustering algorithm based on click stream and user-defined event is proposed.First,a general method is designed to calculate the similarity of the session.Then,an improved DBSCAN algorithm is proposed to get session clustering in the first level clustering.Finally,in the second level of clustering,K-means algorithm is used to cluster users according to the distribution of all sessions generated by each user in different session clustering.(2)A text classification model based on BERT is proposed for text content recognition in user behavior analysis.Knowledge from training data is better introduced into the fine-tuning of BERT model by constructing auxiliary sentences,so as to increase domain related knowledge.At the same time,some different fine-tuning strategies on BERT model are carried out comparative analysis in terms of learning rate,sentence length and model structure.(3)Aiming at the problem of image recognition in user behavior analysis,the clothing image is taken as the object of study.Moreover,a multi feature enhancement module based on attention mechanism is added onto the basis of feature extraction with Res Net-50 structure.It can make up for the lack of spatial details and features lost in the process of upsampling,and enable the resolution of thermal graph used to predict key points of clothing consistent with the input image.(4)Aiming at the problem of video action recognition in user behavior analysis,a neural network model is proposed to solve the problem of weak supervised action location.On the premise that only video-level tags are given,the proposal input and video tags are bridged by the proposal selection layer,so that the video level loss can be migrated to the proposal level loss.Moreover,the network parameters can be effectively fine-tuned,so that the two-level network is successfully trained to locate the action in the untrimmed video.Compared with the research of user behavior based on image recognition,video action recognition is helpful to further recognize the user behavior contained in time sequence images,and improve the accuracy of user behavior analysis.It builds up the foundation for the research of user behavior analysis for massive video.