Stopping Criterion Selection for Efficient Semi-supervised Time Series Classification

Summary

High-quality classifiers generally require significant amount of labeled data. However, in many real-life applications and domains, labeled positive training data are difficult to obtain, while unlabeled data are largely available. To resolve the problem, many researchers have proposed semi-supervised learning methods that can build good classifiers by using only handful of labeled data. However, the main problem of the previous approaches for time series domains is the difficulty in selecting an optimal stopping criterion. This work therefore proposes a novel stopping criterion for semi-supervised time series classification, together with an integration of Dynamic Time Warping distance measure to improve the data selection during a self training. The experimental results show that this method can build a better classifier that achieves higher classification accuracy than the previous approach. In addition, the extended proposed work is shown to have satisfactory result for multi-cluster and multi-class semi-supervised time series classifier.