Research On Group-based Recommendation Via Multi-feature Fusion In Social Networks

With the rapid development of information technology as well as mobile Internet,the development trend of information explosion has become irresistible.Massive information not only brings people more choices,but also makes them often lost in the ocean of information,so that people cannot easily find the information they require.This leads to the problem of information overload.In recent years,the recommendation systems have become an effective tool to solve the problem,and have been successfully developed into typical commercial applications by IT industry giants such as Google,Baidu and Ali.Most of current recommendation systems are designed for individual users.However,as distances among people have been drawn closer and closer,a huge number of online and offline activities in social networks are often carried out in the form of groups,which accordingly produces the demand of recommending information to the groups of users.In recent years,some researchers have begun to pay attention to group-based recommendation systems,and have also put forward some representative solutions as well as ideas.But existing research works still need some improvement.For one thing,social groups possess many forms of existence,including random groups and persistent groups.Recommendation systems also have a variety of scene forms,such as implicit recommendation,explicit recommendation,sequential recommendation.But there still lacks systematic researches on recommendation problem for different types of groups concerning different types of scenes.For another,existing methods are technically simple,most of them just aggregates individual characteristics into group characteristics through different aggregation functions.However,subjective definition of aggregations will introduce errors.Moreover,social groups always possess multiple features(isomerism,dynamics,etc.),so that group characteristics are not the simple superposition of individual characteristics.In order to achieve efficient group-based recommendations,it is necessary to integrate the diversified features in different scenes.In view of this,this dissertation aims at two types of social groups,combines different types of business scenarios,and studies several types of typical group-based recommendation problems.Main research contents and contributions of this dissertation are as follows:(1)This dissertation studies explicit recommendations for random groups.In light of the problem that noise bias will be introduced by existing methods,a group-based recommendation mechanism through multi-dimensional preference distribution is proposed in this dissertation.First,preferences of group members are aggregated into the form of preference distirbutions,in order to represent group preferences from the perspective of characteristic diversification.Second,multi-variate support vector regression model is esatablished to fit the generation of multi-dimensional preference distributions.And iterative weighted least square method is utilized to estimate model parameters.After that,unknown group preference information can be well predicted.Then,a fuzzy multi-criteria decision making scheme is presented to integrate group preference charateristics with the form of multiple dimensions,and to calculate item rankings as recommendation results.Finally,experiments on real-world datasets show that the proposed group-based recommendation mechanism is able to produce more effective recommendation results and has excellent robustness to parameter changes.(2)This dissertation studies implicit recommendations for continuous groups.Aiming at the problem of lacking clear feedback information,a group-based recommendation mechanism through probabilistic reasoning and non-cooperative game is proposed in this dissertation.First,behavioral records are assumed to be decision selection processes influenced by process variables.Second,reasonable priori distributions are assigned for the process of behavioral generation to establish joint probability expression of behavioral generation processes.Based on the observable behavioral results,explicit unknown variables are inferred from the implicit behavioral records with the help of Bayesian posterior probability reasoning.Then,analyzing cooperation and competition mechanism of group members,this dissertation utilizes non-cooperative game model to transform production of recommendation results into the process where group members search for optimal untility through game.Schemes in the state of Nash equilibrium are recommendation results.Finally,experiments on real-world datasets show that the proposed group-based recommendation mechanism is able to produce more effective recommendation results and has excellent robustness to parameter changes.(3)This dissertation studies sequential recommendations for random groups.As data samples contain significant temporal correlation characteristics,this dissertation proposes a group-based recommendation mechanism through Gaussian process regression.First,heterogeneous behavioral records of group members are aggregated into behavioral series of groups.And LDA topic model is employed to extract topic distributions of interactions in each time stamp,as the behavioral charateristics of groups in each time stamp.Second,Gaussian process regression is utilized to model sequential evolution process of group behavioral characteristics and to predict trend of behavioral characteristics in the follow-up time stamps.Then,on the basis of predicted trend of behavioral charateristics,item sets whose characteristics are most close to the predicted ones,are selected as recommendation results.Finally,experiments on three real-world datasets show that the proposed group-based recommendation mechanism is able to produce more effective recommendation results,and is not easily influenced by data sparsity problem.To sum up,this dissertation puts forward some new research ideas for the recommendation problem of random groups and continuous groups in social network,and provides some references for further research on group-based recommendations as well as related topics.