| Résumé | This paper presents a novel approach for the analysis of time series, using Process Mining conformance checking measures as features describing the series’ models, in conjunction with machine learning techniques. The presented approach aims at identifying classes of time series based on their structure content and behavior (normal series, cycles, trends and changes of level). Event logs constructed from time series of different kinds are processed with the Inductive Miner algorithm, producing Petri Nets models. The resulting models are used for creating sets of features for individual time series logs, using trace-fitness measures. These measures are obtained with token-based replay conformance checking. An assessment of the predictive potential of these features is obtained with the Gamma Test, and a collection of supervised machine learning techniques is used for constructing classification models. For some classes, high performance models were found, demonstrating the potential of the proposed approach. However, there were classes for which the Petri Net models produced by the Inductive Miner algorithm did not produce token-based replay conformance checking features with enough predictive power. Overall, Process Mining-based time series classification using conformance checking proved to be a valuable tool. The results obtained are encouraging, but preliminary, and open several avenues to investigate. |
|---|
