IPTV User Behavior Pattern Mining Based On Topic Model

With the development of the Internet,IPTV(Interactive Personality Television)has become more and more widely used in people's lives.Unlike ordinary TVs,users on IPTV can interact with TVs,and users' behaviors are more personalized.Analyze the user's behavioral data and mine the user's behavior pattern can discover the user's viewing preference,carry on the targeted program recommendation,formulate the individualized program menu,can also carry on the targeted advertisement according to the user behavior pattern.The topic model is a statistical model for analyzing large-scale texts.It can map high-dimensional document-word space to low-dimensional document-topic space and topic-word space,thereby realizing document clustering,potential topic mining,etc..The IPTV user's behavior data exists in text form,and the behavior data of each family can be regarded as a document.Therefore,the topic model can be used to mine user behavior patterns from the user behavior data.This article first analyzes the characteristics of the IPTV user's behavior data.The IPTV user is generally a family.The time when different members of the family watch the program is generally different,and the IPTV user's behavior record includes viewing,collecting,and browsing records.Then based on the above features,this paper proposes a Multiple Behavior coupled Time LDA(MT-LDA)model to mine user behavior patterns.Finally,this paper proposes an online learning algorithm for the model.The algorithm can quickly update the model after the new behavior data is generated,and captures the changes in the user behavior patterns and the evolution of the topics.The innovations in this article are as follows:1)This paper proposes an MT-LDA model based on multiple user behaviors.The model improves the cLDA model,integrates the user's viewing,collection and browsing records,and the generated time of these behaviors to mine user behavior patterns..The experimental results show that the model can be used to derive the hidden topics of TV programs,and the distribution of viewing preferences for each family at different time periods can be obtained.Compared to the cLDA model,the topics obtained by the MT-LDA model is more accurately,the MT-LDA model's perplextity is less than the cLDA model.2)This article proposes an online learning version of the MT-LDA model,online MT-LDA.It transfers the data to the model on a time-by-window basis for training.After the new data arrives,it does not need to retrain all the data.It only needs to train the newly added data to updatable model.Online MT-LDA can discover the evolution of topics over time,can more accurately find the changes in program popularity under different topics,and can also discover changes in user behavior patterns over time to provide more accurate personalized services.