A Frequent Pattern Based Time Series Classification Framework
一种基于频繁模式的时间序列分类框架

How to extract and select features from time series are two important topics in time series classification. In this paper, a MNOE (Mining Non-Overlap Episode) algorithm is presented to find non-overlap frequent patterns in time series and these non-overlap frequent patterns are considered as features of the time series. Based on these non-overlap episodes, an EGMAMC (Episode Generated Mixed memory Aggregation Markov Chain) model is presented to describe time series. According to the principle of likelihood ratio test, the connection between the support of episode and whether EGMAMC could describe the time series significantly is induced. Based on the definition of information gain, significant frequent patterns are selected as the features of time series for classification. The experiments on UCI (University of California Irvine) datasets and smart building datasets demonstrate that the classification model trained with selecting significant frequent patterns as features outperforms the one trained without selecting them on precision, recall and F-Measure. The time series classification models can be improved by efficiently extracting and effectively selecting non-overlap frequent patterns as features of time series.