Design of Multiple Classifier Systems for Time Series Data

In previous work, we showed that the use of Multiple Input Representation(MIR) for the classification of time series data provides complementary information that leads to better accuracy. [4]. In this paper, we introduce the Static Minimization-Maximization approach to build Multiple Classifier Systems(MCSs) using MIR. SMM consists of two steps. In the minimization step, a greedy algorithm is employed to iteratively select the classifiers from the knowledge space to minimize the training error of MCSs. In the maximization step, a modified version of Behavior Knowledge Space(BKS), Balanced Behavior Knowledge Space(BBKS), is used to maximize the expected accuracy of the whole system given that the training error is minimized. Several popular techniques including AdaBoost, Bagging and Random Subspace are used as the benchmark to evaluate the proposed approach on four time series data sets. The results obtained from our experiments show that the performance of the proposed approach is effective as well as robust for the classification of time series data. In addition, this approach could be further extended to other applications in our future research.