Mining Segment-Wise Periodic Patterns in Time-Related Databases

Jiawei Han Wan Gong Yiwen Yin Intelligent Database Systems Research Laboratory, School of Computing Science Simon Fraser University, Burnaby, BC, Canada V5A 1S6 E-mail: fhan, wgong, [email protected] Abstract Periodicity search, that is, search for cyclicity in time-related databases, is an interesting data mining problem. Most previous studies have been on nding full-cycle periodicity for all the segments in the selected sequences of the data, that is, if a sequence is periodic, all the points or segments in the period repeat. However, it is often useful to mine segment-wise or point-wise periodicity in time-related data sets. In this study, we integrate data cube and Apriori data mining techniques for mining segment-wise periodicity in regard to a xed length period and show that data cube provides an e cient structure and a convenient way for interactive mining of multiple-level periodicity. Introduction Periodicity search, that is, search for cyclic patterns in time-related data sets, is an important data mining problem with many applications. Most previously studied methods on periodicity pattern search are on mining full-cycle periodicity in the sense that every point in the period contribute to the part of the cycle, such as all the days in the year contribute (approximately) to the season cycles of the year. However, there exists another kind of periodicity, which we call segment-wise periodicity in the sense that only some of the segments in a time sequence have cyclic behavior. For example, Laura may read Vancouver Sun at 7:00 to 7:30 every weekday morning but may do all sorts of things afterwards; Company W's stock may rise almost every Wednesday but could be unpredictable at other time slots (see Figure 1); and Jack may work regularly (full-cycle periodicity) during working hours but he can only be found at 9:00{10:00 every Monday morning (segment-wise periodicity). These examples show that segment-wise periThe research was supported in part by the research grants from the Natural Sciences and Engineering Research Council of Canada, Networks of Centres of Excellent Program of Canada, MPR Teltech Ltd., and B.C. Advanced Systems Institute. Copyright c 1998, American Association for Arti cial Intelligence (www.aaai.org). All rights reserved. Time Week -Two Week -Three Week-Four Week-Four Week-Five Week-Seven Week-Six Week -One Stock Price