Time Series Classification with Max-Correlation and Min-Redundancy Shapelets Transformation
Yu Wei, Libin Jiao, Shenling Wang, Yinfeng Chen, Dalian Liu · 2015
Shapelet is a fragment of a time series that can be used to represent class characteristics of the time series. Its application in time series classification is extensively discussed and many improvements have been made. One of them is to separate the construction of classifier from shapelet selection process which is called shapelet transformation. It is more flexible and provides us with more freedom and possibilities to make it better. First of all, the consumption of time is still a key problem that needs to be improved. Secondly, how to find the perfect shapelets that are most representative with appropriate amount is another important question we need to answer. In this paper, we first reduce the calculation by merging the existing speed-up techniques with bounding of sliding window. Then, we apply Max-Correlation and Min-Redundancy feature selection strategy in shapelet selection process to choose proper shapelets. We introduce ReliefF to find shapelet candidates with strong class correlation, and eliminate the redundant candidates with less ReliefF weights than their similar candidates. By applying our solution in actual datasets, we have proved the feasible of our strategy in actual application.