Research On Personalized Recommendation System

In this era of overloaded information, personalized recommendation technologycan help users find the information they are interested in, push the information that theuser may be interest in to the user. In recent years, not only the study of the personalizedrecommendation system has attracted extensive attention in the academic and businesscommunities, but also the application has achieved good results in practice. However,there are still some problems and challenges for this system, and there is a certaindistance from theoretical research to practice. Based on the actual demand, this thesisanalyzes and studies the overall structure and working principle of the recommendationengine, and focuses on the core recommendation algorithm part. The main work of thisthesis includes the following:1. Generally speaking, the studies of personalized recommendation have focusedon recommendation algorithm, but did not do it from the overall structure of therecommendation system. Through the overall grasping of the various modules of thewhole recommendation system, this thesis build recommendation system frameworkwith a strong practicality. Taking into account the computational complexity of largeamounts of data and respond in real-time, this thesis divide the system into online partsand offline part, the offline part is responsible for calculating large amounts of data andgenerating intermediate results, and online part use these intermediate results to makerecommendations for users. Then, this thesis make a analysis and instruction on eachmodule from the technical requirements and functional achieving aspects, and make afocus analysis on how to solve the cold start problem and how to make a real-timecalculation on the ranking of the recommended results through user feedbackinformation.2. For collaborative filtering algorithms have the weakness of low real-time abilityand scalability, and to improve the prediction accuracy, this thesis presents acollaborative filtering recommendation algorithm based on multi-interest of theK-means clustering, including two aspects: score prediction and TopN recommendation.First, clustering the user based on K-means algorithm, and then the current classifying users into a certain user to be focused, it can calculate and recommend results for usersby Collaborative Filtering recommendation. Considering different users with multipleinterests, similarity calculation formula has been improved. On rating predictionproblem, to solve the prediction accuracy reducing caused by score matrix sparsenessproblems, this thesis first fill out the sparse matrix through the SVD dimensionreduction method, so as to improve the accuracy of grading forecast. Taking intoaccount the influence of time factor on TopN recommendation, this thesis improve theuser similarity calculation formula. The experiment show that compared to traditionalcollaborative filtering recommendation, the improved algorithm can get better results.3. Assumes that the user and the item is a node in the graph model, therecommended process is to find out the item node which has the largest correlation withthe user node, which will not be constrained by the information mining technology.Moreover, the graph model has good scalability, so this thesis research on personalizedrecommendation algorithm base on user-item bipartite graph. Considering that atraditional algorithm, based on graph model, has the very high time complexity, thisthesis make some improvement based on the old algorithm. Then, considering theinfluence of time information on the user’s interests, this thesis introduce the timeinformation into the user-item bipartite graph to be a Period of time users-itemsbipartite graph, and make recommendations on this graph model. Experimental resultsshow that our method can make higher accuracy than non-temporal methods.