A Dimensionality Reduction Technique for Time Series Classification Using Additive Representation

Time series classification has become one of the most prevalent time series mining tasks. There has also been extensively demonstrated in a variety of domains that other non-time series data can also be transformed into time series data. However, time series data are often large and occasionally contain noise and outliers. Therefore, there have been numerous efforts to resolve the problem by reducing the dimensionality of the data or coming up with new data representations. In this work, we propose a very simple approach to reduce the dimensionality of the data and to significantly speed up time series classification without sacrificing classification accuracy. Our experiment results show that our new additive representation could achieve better classification accuracy in most datasets when compared to the popular SAX symbolic representation and also has better accuracy in more than half of the datasets when compared with the full dataset classification. Our proposed representation also has higher average compression ratio that of SAX, giving our representation a competitive choice when dimensionality reduction is needed.