Authors: Omar Bahri (Utah State University), Pouya Hosseinzadeh (Utah State University), Soukaina Filali Boubrahimi (Utah State University), and Shah Muhammad Hamdi (New Mexico State University)

The amount of time series data being collected has exploded in parallel with the usage of sensing devices and the improvement of storage capacities. As a result, massive time series datasets have been produced in recent years. In order to analyze these datasets, substantial efforts are being invested in the implementation of temporal rule mining algorithms. Coupled with the knowledge of domain experts, these approaches are extremely helpful for interpreting time series data and reaching important conclusions on the underlying systems. On the other hand, numerous algorithms have been proposed for the time series classification task. In this work, we aim to increase the interpretability of current time series classification methods, by filling the gap between rule discovery and classification in time series mining. We propose rule transform (RT), an algorithm that uses temporal algebra to transform a time series dataset into a new feature space consisting of the support of the most prominent temporal rules. Thus, the generated feature space can be 1) qualitatively interpreted by domain experts and 2) used for classification. We evaluate our algorithm on the UEA archive, and prove that it produces accuracies superior to state-of-the-art time series classification algorithms with the additional interpretability edge. Then, we apply it to a real-life solar flare dataset and discuss our findings.