Research On Some Key Problems Of Web-based Recommendation Systems

With the popularity of the Internet and the rapid development of network technology, the amount of information on the Internet gradually showed the trend of information overload. In addition, with the rapid development of e-commerce, the e-commerce size was expanding continually and the quantity and variety of goods on e-commerce site were growing rapidly. The rapid development of the network technology brought convenience for the users. At the same time, it makes user difficult to make choices when facing massive messages. It was difficult for users to collect, organize, analyze information and make decisions by themselves. Therefore, recommendation systems were produced. In essence, the recommendation systems were a kind of information filtering technology. How to effectively analyze information and users'characteristics, analyze the collected historical data and produce effective decision are the main problems in recommendation systems. It also raises many challenging issues to computer science and engineering practice.In this paper, some key problems of Web-based recommendation systems were researched. The main contents include:methods and technologies that can improve recommendation accuracy, current recommendation evaluation indexes, session recommendation diversity index and how to use the users'context information to improve the recommendation diversity during users'session. The major studies of this dissertation include:(1) For improving the recommendation accuracy, a new collaborative filtering algorithm combining dynamic k-Nearest-Neighborhood and SlopeOne is proposed. SlopeOne algorithm uses linear regression model to solve data sparisity problem. k-Nearest-Neighborhood method based on users'similarities can optimize the quality of ratings made by users participating in prediction. Firstly, different numbers of neighbors for each user are dynamically selected according to the similarities with other users. Secondly, average deviations between pairs of relevant items are generated on the basis of rankings from neighbor users. At last, the object ratings are predicted by linear regression model. Experiments on the MovieLens dataset show that the proposed algorithm gives higher prediction accuracy and is more robust to data sparsity than SlopeOne. It also outperforms other collaborative filtering algorithms on prediction accuracy.(2) For the evaluation indexes of recommendation systems, we found that diversity was increasingly becoming one of the important measures of recommendation systems. But current diversity indexes cannot indicate the recommendation diversity during users' session and the existing methods for improving recommendation diversity always evaluate recommendation diversity only by single recommendation. In Web-based recommendation system application, the fact was ignored that the actual unit of users' behavior was session. A kind of evaluation index about session recommendation diversity is proposed. It evaluates the ratio of all recommended items without repetition to the total number of recommendation positions during user's session. Session recommendation diversity is able to evaluate how many items are browsed by a user during the session and it can explore the long-tail products and improve the conversion rate of recommendation.(3) We made a recommendation diversity survey of the state-of-the-art and found that existing methods for improving recommendation diversity always came at the expense of precision and could not evaluate the overall recommendation diversity during user's whole session. Traditional recommendation systems had poor session recommendation diversity because there were too many repeated nodes in the session recommendation trees. An offline-UICF session recommendation algorithm is proposed to eliminate the recommendation redundancy. It uses user's session as a whole unit to recommend. Experiments on MovieLens dataset show that offline-UICF algorithm has substantially higher session recommendation diversity and better recommendation precision.(4) For utilizing user's context information to improve the recommendation quality, we take the user's browsing path during the session as context information to improve the user's session recommendation diversity. An online-UICF session recommendation algorithm is proposed. By creating session recommendation lists for each active user, the algorithm can avoid recommendation loops or weak recommendation loops in the recommendation trees. Experiments on MovieLens dataset show that online-UICF algorithm has both higher session recommendation diversity and better recommendation precision.Based on the research of recommendation accuracy and diversity, we design and implement a Web-based session recommendation system named SessionRecommender. SessionRecommender utilizes SlopeOne collaborative filtering recommendation algorithm based on dynamic k-Nearest-Neighborhood, offline-UICF and online-UICF algorithms to recommend. It can provide high accuracy and diversity. In SessionRecommender, other exist recommender engines can be added to our session recommendation modules. SessionRecommender provides a new tool for Web-based application to apply recommendation systems.In summary, this dissertation is dedicated to some key problems related to Web-based recommendation systems, such as accuracy of prediction, recommendation evaluation indexes and session recommendation diversity. We propose some algorithms to improve recommendation accuracy and session recommendation diversity. Lots of theoretical analysis and experiments show that these approaches are efficient and effective. We hope that these approaches and techniques could make contributions to Web-based recommendation systems.