Research On Data-driven Prediction Algorithm Of IPTV User Engagement

The Internet and the Internet of Things have influenced the people's life.Internet Protocol Television(IPTV)has greatly changed the way of watching videos.Under such condition,people have a growing demand for quality of experience(QoE)in IPTV services.Traditional scoring evaluation methods can't meet the demands of video content providers and network operators regarding user satisfaction.Instead,collected mass data are used to evaluate user's QoE.Therefore,how to measure user's QoE with subjective characteristics objectively and effectively has become a hotspot and difficulty in current research.Based on this,in this paper,we conduct research by focusing on user engagement that is closely related to the user's QoE and selecting the user viewing time with objective characteristics.First of all,we preprocess and select features of the data collected by the IPTV set-top box,and then design related models and algorithms to model and predict the user viewing time,so as to achieve data-driven prediction of IPTV user engagement.The main work of the thesis is reflected as the following three aspects:First,preprocessing and feature selection are performed on the data collected by the IPTV set-top box.For the missing and duplicated data records,they are integrated and cleaned.In addition,feature selection is performed on the collected data.Concretely,combining the Relief algorithm with clustering,a hybrid feature selection algorithm is proposed.Theoretical analysis and simulation results show that this method can effectively extract the most valuable feature attributes for the engagement prediction.Then,on the basis of feature selection,a prediction model of user viewing time that can reflect user engagement is established.In particular,traditional k-Nearest Neighbor(KNN)regression algorithm is combined with the classification and regression tree(CART),then a prediction model of user viewing time based on weighted KNN-CART is designed.Compared with the existing algorithms,the new model effectively shortens testing time under the premise of ensuring prediction accuracy.Finally,based on the improved extreme learning machine(ELM),another prediction model of user viewing time is established.Different from the traditional prediction method,the ELM can synchronously complete feature selection and modeling.Based on Agglomerative Nesting(AGNES)feature clustering,an improved ELM modeling and prediction algorithm is proposed.The model structure is simplified on the premise of ensuring accuracy.