Research On Recommendation Based On Local Modeling And Global Fusion

With the rapid development of Internet,personalized recommendation system has become an important tool to solve the problem of information overload.In recent years,the number of users and items on the network have grown rapidly,which makes it increasingly difficult for global recommendation models to make accurate recommendations.These global recommendation models will not only face greater overhead in training and storage,but also face the sparsity and noise caused by large-scale data.This thesis studies how to split the global data set into several local groups,each composing a subset of all available users and items,how to conduct local modeling in each local group,and how to fuse the local results for global recommendation result.Based on the idea of local modeling and global fusion,this thesis conducts in-depth research on the tasks of rating prediction and Top-K recommendation for recommendation systems,and proposes the following three models:(1)A local matrix approximation method based on feature divergence measure for rating prediction(FDLMA).FDLMA constructs local user-item matrices by first selecting some(user,item)pairs as anchors and then finding the neighbors for each anchor.To be specific,FDLMA proposes a representative degree based anchor selection method to select some user-item pairs with high density and far apart from each other as anchors.Besides,FDLMA proposes to use feature divergence measure to compute the asymmetric distances in between users and items,thus constructing local matrices with more concentrated interests.(2)A destructure-and-restructure matrix approximation framework for rating prediction(DRMA),which approximates user-item rating matrix through three stages:matrix destructure,local matrix approximation and matrix restructure.In the matrix destructure stage,to avoid the problem of first pre-training the feature vectors of users and items and then calculating the distance for local groups construction,DRMA has proposed a two-stage random walk method on the user-item bipartite graph to measure the distance in between users(items).In the stage of local matrix approximation and matrix restructure,DRMA proposes to implement the differentiated local training and fusion of local results according to the data distribution of different local matrices.(3)A local ranking and global fusion method(LRGF)for Top-K recommendation.We first propose a pairwise gap-aware Bayesian personalized ranking model(pgBPR)to integrate preference gap between a pair of items into ranking.In the local ranking stage,we exploit pgBPR to generate local ranking lists for users in each local group.In the global fusion stage,a group-aware and order-scoring decision fusion strategy is proposed to fuse the local lists for global recommendation lists.We conduct experiments on real-word datasets to vertify the effectiveness of our proposed three models,and compare them with other competitive recommendation models.Experiments show that the proposed three models can achieve very good recommendation performance.FDLMA and DRMA achieve lower prediction errors not only than global matrix factorization models,but also local matrix approximation approaches.LRGF outperforms other recommendation models in terms of ranking metrics and hit ratio in most cases.