Fusion of similarity measures for time series classification
HAIS'11 Proceedings of the 6th international conference on Hybrid artificial intelligent systems - Volume Part II
INSIGHT: efficient and effective instance selection for time-series classification
PAKDD'11 Proceedings of the 15th Pacific-Asia conference on Advances in knowledge discovery and data mining - Volume Part II
| Hi-index | 0.00 |
Time-series classification is an active research topic in machine learning, as it finds applications in numerous domains. The k-NN classifier, based on the discrete time warping (DTW) distance, had been shown to be competitive to many state-of-the art time-series classification methods. Nevertheless, due to the complexity of time-series data sets, our investigation demonstrates that a single, global choice for k (= 1) can become sub optimal, because each individual region of a data set may require a different k value. In this paper, we proposed a novel individualized error prediction (IEP) mechanism that considers a range of k-NN classifiers (for different k values) and uses secondary regression models that predict the error of each such classifier. This permits to perform k-NN time-series classification in a more fine grained fashion that adapts to the varying characteristics among different regions by avoiding the restriction of a single value of k. Our experimental evaluation, using a large collection of real time-series data, indicates that the proposed method is more robust and compares favorably against two examined baselines by resulting in significant reduction in the classification error.