Research On Personalized Recommendation System Based On Collaborative Filtering

With the advent of the information era,how to find useful information in the vast amounts of information is a challenge.The recommendation system is one of the tools to deal with this challenge and is focused by more and more researchers.As a hot branch of the recommendation system,collaborative filtering(CF)analyzes users' behaviors instead of the contents of the recommended objects.A major advantage of CF is that it can be applied in any recommendation fields.The core of CF is to calculate the similarity among the users or recommended objects.Therefore,the accuracy of the similarity calculation determines the quality of the recommendation results.This thesis focuses how to improve the accuracy of similarity calculation.The main contentes of this thesis are organized as follows:1)Make a futher study related knowledge of CF recommendation system,including the principle of recommendation,the steps of implementation,the classification of recommendation and so on.It introduces two kinds of neighbourhood collaborative filtering in detail.They are a User-based CF and an Item-Based CF.And the computation cost,application scene and real-time performance are compared synthetically.Also,it summaries the traditional calculation of the similarity and the problems of cosine similarity,pearson correlation coefficient and jaccard coefficient with the instances in this thesis.2)Then,it makes data cleaning,transformation and feature analysis for the original data.Based on the characteristics of the data and the problems of the traditional similarity method,it explores the influencing factors,and proposes a similarity calculation method based on multiple weight.The method combines the frequency items,active users and time information into the jaccard coefficient,thus making the comprehensive measure of similarities between items.The experiments show that similarity calculation method based on multiple weight can effectively improve the accuracy of CF.3)In the CF,the number of users' neighbors will indirectly affect the prediction.It introduces the PSO and then uses the method to determine the number of neighbors.The method aims to improve the accuracy of CF by reducing the MAE effectively and improve the computational efficiency by determining the number of neighbor with fewer iterations.4)Combining with the proposed method,it designs and implements the prototype of personalized recommendation system.The system prototype is designed as the off-line and on-line platforms.The off-line platform is mainly responsible for the heavy works of calculating similarity,confirming the number of neighbors and generating the list of recommendations.The on-line platform is mainly responsible for the light weight works of calculating users' behaviors and checking the users' recommendation lists,and so on.