The Comparison of Traditional Collaborative Filtering Methods and LDA-variant on Diversity, Novelty and Correlation for Recommendation Task

• Main Authors: CHUANG, CHENG-WEI, 莊証幃
• Other Authors: HSIAO, WEN-FENG
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/87z9n8

碩士 === 國立屏東大學 === 資訊管理學系碩士班 === 105 === Past research on recommender systems mostly focused on improving the recommendation accuracy. Although accuracy is the key factor for the success of a recommender system, the accuracy alone is not enough to evaluate the practical utility of a recommender system. Therefore, in recent years, many measures different from accuracy have been proposed, such as diversity, novelty, coverage, serendipity and unexpectedness to compensate the accuracy measure. In this paper, we used Apache Mahout to generate the collaborative filtering recommendation. We also employed the JGibbLDA tool to construct the LDA topic model to infer the topic probability distribution of the items or the users, and further by calculating the similarities to get the LDA based recommendation. In this paper we first filtered out the indicators for diversity and novelty that currently get many citations. We then compared the performance of the recommended methods on these indicators, and suggested the corresponding recommended methods for the users that can balance accuracy, diversity, and novelty, in order to increase users’ satisfaction.According to the experimental results of the film dataset, it is found that the cooperative filtering with LDA can obtain a high correlation, and the traditional cooperative filtering can obtain higher diversity. The project based CF and the LD based on the project can get more High novelty. The experimental results of the joke dataset found that the project-based CF was highly correlated with LDA-based synergistic filtering available for higher diversity, project-based CF and LDA-based CF-based CF can get a higher novelty.