Classification system for time series data based on feature pattern extraction
Hiroshi Sugimura, Kazunori Matsumoto · 2011
This paper proposes a system which acquires feature patterns and makes classifiers for time series data without using background knowledge given by a user. Time series data are widely appeared in finance, medical research, industrial sensors, etc. The system acquires the feature patterns that characterize similar data in database. We focus on two aspects of the feature pattern: global and local frequency. Our purpose is to acquire features of each data by extracting these patterns. The system cut out subsequences from time series data. Several representative sequences are extracted from these subsequences by using clustering. Feature patterns are acquired from these representative sequences. For this purpose, we develop a method that applies TF*IDF weight technique, which is often used in text mining, to time series data. The time series data are classified by using the acquired feature patterns. In accordance with a criterion that is based on the entropy theory, feature patterns are improved by the automatic process, generation by generation, using the genetic algorithm. By using the final and optimized feature patterns, we build a decision tree that determines future behaviors. We explain how these two tools are combinatory applied in the entire knowledge discovery process.