Tabu search for dynamic time warping global constraint learning

Measuring similarity or distance between two data points is fundamental to many Machine Learning algorithms such as K-Nearest-Neighbor, Clustering etc. Depending on the nature of the data point, various measurements can be used. DTW is largely used for mining time series but it is not adopted to large data sets because of its quadratic complexity. Global constraints narrow the search path in the matrix which results in a significant decrease in the number of performed calculations. The distance between examples from the same class is small. Instances from different classes are with large distances. A field called metric learning is introduced to make such criteria. In some time series classification tasks, it is a common case that two time series are out of phase, even they share the same class label. An appropriate constraint of DTW can strongly improve the classification performance. It is to choose the appropriate size of the global constraint. A Tabu search algorithm is used to find the optimal size of the global constraint. Results show the efficiency of the proposed method in terms of the improvement of the classification results and the CPU time.