Research On Information Recommendation Technology Under The Background Of Big Data

To relieve the three inherent problems of data sparsity,cold start and scalability in the recommender systems,related recommendation algorithms are presented.The exper-imental results show that all the proposed algorithms can achieve better recommendation performance.Firstly,a personalized recommendation algorithm based on weighted bipartite network be called INBIw(Improved weighted Network-Based Inference)is presented.By using the principles of heat conduction and resource allocation in physics,recommender system is modeled as a weighted bipartite network,weight of each edge is the rating of relevant us-er for item,and the rating is regarded as resource in resource allocation.A free tunable parameter ? is introduced to depress the influence of high degree node.With two step-s of resource allocation process that from item to user,and then from user back to item for personal recommendation.In order to evaluate the recommendation performance of INBIw,ranking position rate(RPR)and hitting rate(HR)are calculated.The results of ex-periment based on MovieLens data set show that the INBIw outperforms previous methods,including the Global Ranking Method(GRM),Collaborative Filtering(CF),Network-Based Inference(NBI),weighted Network-Based Inference(NBIw),and Latent Interest and Topic Mining(LITM)with respect to RPR and HR.Specifically,it performs well and gives a more accurate prediction.After further analysis,we discovered that the recommendation results of INBIw are insensitive to the amount of data and length of the recommendation list.Thus,INBIw can deal with data sparsity and is able to satisfy the varied requirements of real sit-uations.Meanwhile,the experimental results on several large data sets also confirmed that effectiveness,scalability and Robustness of the proposed recommendation algorithm.Next,a personalized recommendation algorithm be called ITNBIw(Improved weighted Tripartite Network-Based Inference)is presented.By using the principles of heat conduction and resource allocation in physics,recommender system is modeled as a weighted tripartite network of user-item-item's property.The resource allocation process consists of two steps:first from item to user then back from user to item,and first from item to item's property then back from item's property to item,respectively.And then to combine with them.While computing the integrated resource of item,a scale parameter A is introduced.In order to evaluate the recommendation performance of ITNBIw,RPR and HR are calculated.At first,to determine the optimal ? according to the values of RPR and HR.Then,to compare the performance of proposed algorithm and other algorithms in the case of ?opt.The results of experiment based on MovieLens data set show that the ITNBIw outperforms previous methods,including the GRM,CF,NBI,NBIw,and LITM with respect to RPR and HR.The further experimental results show that,the best case is introducing edge weights and free tunable parameter ? in tripartite network meantime,next is introducing edge weights and no free tunable parameter ?,and the worst case is neither introducing edge weights nor free tunable parameter?.Finally,a personalized recommendation algorithm be called CFBDF(Collaborative Filtering Based on Data Filling)is proposed.The presented recommendation algorithm applies three data filling approaches,two recommendation strategies and one user similarity computation for collaborative filtering recommendation.Three data filling methods for non-rating data in rating matrix are:(1)Filling data using weighed average of row and column ratings;(2)Filling data using mode average of row and column ratings;(3)Filling data using median average of row and column ratings.Two recommendation strategies are:(1)Taking filling data for predicative rating directly;(2)Seting the rating matrix filled data as a pseudo rating matrix for collaborative filtering recommendation.One user similarity computation method is based on weighted bipartite network and resource allocation principle.The key is to calculate the asymmetric user resource allocation weighted matrix and translate it into a symmetric user similarity matrix.The experimental results on the MovieLens data set show that all these recommendation strategies can effectively to alleviate the trouble of rating data sparseness and can gain better recommendation accuracy.