Entropy-based Symbolic Aggregate Approximation Representation Method for Time Series

Symbolic Aggregate approXimation (SAX) is one of the most common dimensionality reduction approaches for time-series and has been widely employed in lots of domains, including motif discovery, time-series classification, and fast shapelets discovery. However, SAX only considers the average value of the segment but ignores other essential features. As a result, two segments with different shapes but similar average values are transformed into the same symbol. In this paper, we first propose Entropy-based Symbolic Aggregate approXimation (EN_SAX) to overcome this drawback. The EN_SAX improves the original SAX by capturing an additional characteristic in a segment by using the time-series entropy. Then we present a modified similarity measurement to compute the similarity between pairwise time-series. We extensively compare EN_SAX with state-of-the-art improved SAX approaches in time-series classification framework using ten real-life datasets to demonstrate the effectiveness of the proposal. Experimental results show that EN_SAX outperforms other approaches in classification accuracy.