Enriching Multivariate Temporal Patterns with Context Information to Support Classification

In this paper we consider classification tasks where the class depends on the co-evolution of multiple variables over time, for instance, “if A happens before B and in the meantime we do not observe C, then we have a failure of class X”. We present a two-phased approach to derive such patterns from data. In the first step, we seek the most specific pattern that still matches all data from one class and in the second step we constrain the pattern further, such that it discriminates with respect to other classes. While the second step is directly motivated by the classification task, the first step enables the user to better match his or her mental model of the temporal process to the patterns derived by the classifier. The experimental evaluation on the libras dataset has shown that the additional first step not only improves the interpretability, but also the classification results.