Design And Implementation Of Recommendation Method Based On Implicit Feedback For Video Conference System

Modern netizens often get lost in the massive content of the Internet,and do not know how to find content that they might be interested in.Personalized recommendation system matches users and content,assists users in making decisions,and satisfies users' information needs.It plays an important role in the Internet+ era and has expanded into more fields like video conferencing system.Video conferencing system provides remote real-time communication services via network audio and video,reduces the communication cost,and begins to try to penetrate from the institutional users to individual users,aiming to provide a platform for communication.Recommendation technology can also be applied to video conferencing system to help users find the activities they like and want to participate in.It can be noticed that in most systems,including video conferencing systems,user-item rating feedback are not available and only implicit feedback such as clicks and conference participation can be collected.Recommendation system can only model user preferences with implicit feedback.This thesis attempts to focus on the implicit feedback recommendation in the video conferencing field,analyzes the data pattern of the video conferencing system,and designs the recommendation algorithm in a targeted manner.The Sim BPR personalized ranking recommendation algorithm that combines the information of the item content and implicit feedback is proposed.Based on this,we construct a recommendation module for the video conferencing system.This thesis first analyzes the history and development trend of recommendation system and video conference system,and then describes the existing implicit feedback recommendation methods,which can be divided into three types: one-class collaborative filtering,recommendation with auxiliary information and learning-to-rank.By observing the practice of the industry,it is clear that the method based on pairwise learning-to-rank is the best solution.This thesis researched the Bayesian Personalized Ranking(BPR),which maximize the posterior probability of model parameters to perduce recommendation.A method for implementing BPR on a distributed system is then proposed.In view of the fact that the item content information has always been an effective means of recommendation,this thesis uses the item description text to represent the item,and designs a three-in-one similarity calculation method based on semantic similarity,keyword similarity and category similarity.This thesis then introduces a further preference hypothesis,arguing that users will be more interested in items that are more similar to the items they have interacted with in the past,thereby classifying the set of items that the user has not interacted with into a high-similarity items set,which called false feedback and a low-similarity items set,called missing feedback.Based on this,this thesis improved the BPR algorithm,and proposed the Sim BPR algorithm that including true feedback samples and false feedback samples.In addition,the random sampling method of the BPR algorithm has a optimization space,so in this thesis,the positive and negative samples are sampled based on time and popularity.It is observed that the industry popularizes the use of large-scale context information to model recommendations.This thesis also designs the context features that can be collected for the video conferencing system,including the number of participants of a conference and the friendship between participating users,and integrates them into recommendation model,then a new user prefernece predictor was built for Sim BPR.Then the thesis design and built a recommended modules for video conferencing systems based on Spark,Hive,HBase and other distributed technologies.In the module design,the modules are divided into offline,near-line and online sub-modules,which are responsible for non-real-time,quasi-real-time and real-time computing tasks.In the system,two data collect methods,full upload and log upload,are designed,then use the star model in data warehouse for data modeling.Finally,this thesis details the design and implementation of each sub-module.In the third part,the thesis made an offline experiment on the Sim BPR algorithm,and selected two advanced methods: BPR and Ao BPR as the comparison algorithn.The recommended list length was set as 5,and used two data sets for the test.The result indicates that Sim BPR is better than the comparison algorithn in AUC,MAP and NDCG indicators,which further verified the positive effect of use of item information in recommendation.The experiment also tested the effect of different values of sub-similarity weights in the item similarity calculation and false feedback weights in the model.