Mining Frequent Patterns from Dynamic Data Streams Using Support Approximation

• Main Authors: Chao-Wei Li, 李兆偉
• Other Authors: Kuen-Fang Jea
• Format: Others
• Language: en_US
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/74852547458965418836

博士 === 國立中興大學 === 資訊科學與工程學系所 === 101 === In many real-life applications nowadays, the data are present in the form of continuous streams, which has made information retrieval and knowledge discovery more difficult and challenging. In this dissertation, we study the practical problem of discovering frequent patterns (large itemsets) from streaming data, namely data-stream frequent pattern mining. We also tackle some important issues concerning data stream mining, which include data overloads and concept drifts. Our methodology to the studied problem is called support approximation. It obtains the frequency counts of itemsets by fast estimation instead of enumeration and counting. The basis of support approximation is the frequency relationships between itemsets and itemsets. We have proposed approaches of modeling the frequency relationships in order to construct support approximation functions. As to the issue of data overload, we devise two mechanisms with respective ideas for lessening the workload of the mining system. With regard to the issue of concept drift, we deal with it by means of concept description and concept learning, which are related to the principle of support approximation. According to our experimental evaluations, the proposed method could have high efficiency in streaming data processing, and it could achieve reasonable mining accuracy. Besides, to a certain extent, the issues of concept drift and data overload could be addressed adequately.