Impact of Pattern Dependency on Web Mining Models

• Main Authors: Lin, Yung-Chang, 林永昌
• Other Authors: Wu, Sheng-Tang
• Format: Others
• Language: zh-TW
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/29985312664426971911

碩士 === 亞洲大學 === 資訊多媒體應用學系 === 102 === With the rapid development of World Wide Web and Information Technology, the internet has become an indispensable element in people’s life. Therefore, how to retrieve information from large amounts of data and convert into useful knowledge is a very important issue. It is one of main research works in Knowledge Discovery field as well. Web Mining can be divided into three categories, including Web Content Mining, Web Structure Mining, and Web Usage Mining. Among them, the problem of Web Content Mining is how to accurately and quickly respond to users' needs and to avoid providing too much irrelevant information. As a result, Pattern which is used for describing semantic in a Web Mining system seems to be significantly important. This research work focus on the dependence between patterns and features. We analyze feature selection mechanism for understanding the impact on pattern generation by using Pattern Taxonomy Model. The experimental results show that the ratio method for determining the number of features performs well and also not affected by the varying number of training samples. In addition, among 3 feature selection methods including TFIDF, Information Gain (IG) and Mutual Information (MI), IG is the best method both in two ways of feature number determination mechanisms. It is recommended to use IG as feature selection method in PTM models and use top 70% of features for pattern composition for the best performance.