Research On Personalized Recommendation System Based On Collaborative Filtering And Latent Factor Model

With the rapid development of the Internet industry,the network is full of information.To solve the negative effects of information overload and improve users' utilization of information,an intelligent information filtering system comes into being,which is called personalized recommendation System.The neighborhood-based collaborative filtering technology is a kind of popular algorithm,it analyzes the similarity between user and user,item and item,which is based the statistical principle.Since users who share similar interests tend to like the same items,the system will recommend the corresponding items to the user.Another recommendation algorithm which is based on the latent semantic model,tries to figure out the user's latent interest preference and the item's latent characteristic through machine learning,establish the relationship between the user and the item and then produce the recommendation accordingly.This thesis studies the recommendation system theory and implementation method based on collaborative filtering model and latent semantic model,and compares the differences between the two algorithms.After in-depth research and analysis,this thesis proposes a hybrid solution named latent item feature based collaborative filtering algorithm(LICF)which integrate these two algorithms into one.The algorithm firstly uses gradient descent method to generate user's potential preference vector set and items' implicit feature vector set;then uses item's implicit feature vector set to calculate similarity between items to make a soft classification of items;after that we use users' preference vector and users' history scores,respectively,multi items' implicit vector to generate the possible scores that users may put on these items.Finally,after weighted processing,TOP-N recommendation results will be generated.The LICF generates a better list of recommendations since it overcomes the shortcomings of the neighborhoodbased collaborative filtering algorithm,which only analyze the explicit features of users.The user's potential preference vector can represent the user's interest more granular and the established user model is more precise.Meanwhile,the classification of the item by the item implicit feature vector is more scientific.Secondly,the algorithm alleviates the shortcomings of neighborhood-based collaborative filtering algorithm in the face of sparse data sets.Finally,the recommendation has a reasonable explanation by incorporating the idea of collaborative filtering.Through the three kinds of experiments on MovieLens dataset,the LICF algorithm has a certain improvement in recommendation accuracy compared with the traditional collaborative filtering algorithm.With the increase of the number of implied factors,the MAE of LICF algorithm is significantly lower than the traditional collaborative filtering one's under the same neighbors.Also,with the fixed hidden factors and the increase of the neighborhood numbers,the MAE of LICF-A is much lower than the traditional one's.Finally,with fixed implied factors and neighbors,with the increase of data sparsity,the MAE of LICF-A algorithm is also significantly lower than the traditional neighborhood-based collaborative filtering algorithm which means this problem has been alleviated.