Application And Optimization Of Collaborative Filtering In Video Recommendation

With the rapid development of network transmission technologies and the proliferation of mobile devices,video traffic grows exponentially.This video big data has brought us many opportunities.In addition to traditional video portals(such as Youku and NetFlix),various types of UGC(user generated content)platforms have emerged(such as YouTube,users as both content consumers and providers).Recommender as a bridge connecting video and users has become the core competitiveness of video companies.However,it also poses new challenges to the recommendation system for improving personalized service capabilities from extracting effective information in he video big data.Recent years,collaborative filtering as a content-free personalized recommendation technology has been widely studied in academics and industry.The idea behind it is that users with similar preferences in the past still have similar preferences in the future.Meanwhile,compared to the content-based recommendation,it does not require labor and financial resources to manually collect user information,mark video features because of its content-free characteristic.As the main evaluation metrics,the accuracy and computational complexity is the main research topics in the area of collaborative filtering.The accuracy is important for describing user preferences and thus for accurate recommendation.The computational complexity determines its practicability and quick implementation.As far as collaborative filtering is concerned,the accuracy and complexity of the algorithm is usually a contradiction.How to achieve a trade-off between the two is a technical difficulty it faces.On the basis of analyzing related work of collaborative filtering,this text focuses on the collaborative filtering of videos based on implicit feedback from two aspects:recommendation accuracy and computational complexity.In the aspect of improving the recommendation accuracy,by analyzing the dynamic features of the user preferences over time,this text proposes a video collaborative filtering ranking model based on time information,thereby improving the limitations of the static collaborative filtering model that lacks the ability for modeling dynamic user preferences.In terms of training complexity,this text analyzes the characteristics of optimization object and learning algorithms in existing collaborative filtering model,and theoretically analyzes the reasons for the existence of invalid optimization of the existing learning algorithms:gradient " offset”.Based on this analysis,this text proposes to improve the sampling method to alleviate the problem of slow learning caused by it.The main work of this text is as follows:1.By analyzing the characteristics of user preferences changing over time,as long-term evolution,local stability,this text proposes a collaborative filtering model based on user local preferences similarity stable characteristic.This implicit time model not only controls the computational complexity of the model but also improves prediction accuracy.At the same time,in order to prevent the model from overfitting,this text proposes to integrate the Bayesian personal sorting model into the proposed time model.Based on this integrated model,this paper proposes an approximate optimization algorithm with two-stage learning.2.Based on the analysis of the optimization target and corresponding learning methods of stochastic gradient descent in the Bayesian personalized collaborative filtering recommendation model,this text presents the mathematical form of the reason why gradient "offset" occurs in the learning process of this algorithm.Based on this form,this text proposes a user clustering based video popularity sampling algorithm to mitigate the inefficient optimization caused by the gradient "offset",thereby improving the efficiency of model learning.