Authors:
Ajumobi Udechukwu
;
Ken Barker
and
Reda Alhajj
Affiliation:
University of Calgary, Canada
Keyword(s):
Trend analysis, Time-series mining, Knowledge discovery and data mining
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Artificial Intelligence and Decision Support Systems
;
Biomedical Engineering
;
Business Analytics
;
Data Engineering
;
Data Mining
;
Databases and Information Systems Integration
;
Datamining
;
Enterprise Information Systems
;
Health Information Systems
;
Sensor Networks
;
Signal Processing
;
Soft Computing
;
Strategic Decision Support Systems
Abstract:
Trend analysis has applications in several domains including: stock market predictions, environmental trend analysis, sales analysis, etc. Temporal trend analysis is possible when the source data (either business or scientific) is collected with time stamps, or with time-related ordering. These time stamps (or orderings) are the core data points for time sequences, as they constitute time series or temporal data. Trends in these time series, when properly analyzed, lead to an understanding of the general behavior of the series so it is possible to more thoroughly understand dynamic behaviors found in data. This analysis provides a foundation for discovering pattern associations within the time series through mining. Furthermore, this foundation is necessary for the more insightful analysis that can only be achieved by comparing different time series found in the source data. Previous works on mining temporal trends attempt to efficiently discover patterns by optimizing discovery proc
esses in a single run over the data. The algorithms generally rely on user-specified time frames (or time windows) that guide the trend searches. Recent experience with data mining clearly indicates that the process is inherently iterative, with no guarantees that the best results are achieved in the first run. If the existing approaches are used for iterative analysis, the same heavy weight process would be re-run on the data (with varying time windows) in the hope that new discoveries will be made on subsequent iterations. Unfortunately, this heavy weight re-execution and processing of the data is expensive. In this work we present a framework in which all the frequent trends in the time series are computed in a single run (in linear time), thus eliminating expensive re-computations in subsequent iterations. We also demonstrate that trend associations within the time series or with related time series can be found.
(More)