Time-Sensitive Stream Pattern Mining under Bounded Memory Constraints

• Main Authors: Chien-Chan Lee, 李建誠
• Other Authors: Ming-Yen Lin
• Format: Others
• Language: en_US
• Published: 2011
• Online Access: http://ndltd.ncl.edu.tw/handle/15010926316533204469

碩士 === 逢甲大學 === 資訊工程所 === 99 === Mining frequent itemsets over data streams is a challenging problem in recent year. The mining in the sliding window model is more desirable since people generally have more interest in the change of recent data. The transaction-sensitive sliding window model considers the window of fixed numbered transactions. In contrast, the time-sensitive sliding window model handles the data stream during a user-defined time interval and is more practical. Most time-sensitive sliding window algorithms assume that the memory is always available and enough for holding the entire transaction data in a window. Nevertheless, these algorithms would fail if bulk data comes in a short burst for the latest sliding window, a common scenario for a particular event in data stream applications. In this thesis, we propose an algorithm, called Tisbom, to solve the problem of mining frequent patterns in time-sensitive stream data using a bounded memory. The Tisbom algorithm uses a condensed representation for each block for efficient window sliding. A block allocation mechanism is invoked to re-adjust the block sizes in the window when the memory is insufficient for the new incoming block. The Tisbom algorithm is characterized in two phases, approximation phase and mining phase. The approximation is to merge similar transactions controlled by a difference threshold. The mining phase is invoked when a user submits a query with a specified support threshold for the frequent patterns in the sliding window. The experiment results show that Tisbom is efficient and has good precisions and recalls for mining time-sensitive stream using a bounded memory.