Predictive temporal patterns discovery.

In recent years, the use of frequent temporal patterns as features for classification has increasingly been used and investigated. In this process, commonly frequent patterns are mined from each class separately. Then the patterns are unified, and feature selection methods may be employed, which are given to induce a classifier. However, this approach is very time consuming since the mining of each class separately takes time. In this paper, we introduce the Saraswati suite that can modify a temporal patterns discovery algorithm into a predictive temporal patterns discovery algorithm, which we demonstrate on Time Intervals Related Patterns. The suite enables predictive patterns to be favored in runtime, while mining both classes simultaneously to discover these patterns. This is through the use of a novel stopping criteria that we call the Saraswati selection criteria and strategies suite. Since the selection criteria are based on the patterns' metrics, such as their frequency in each class or their reoccurrence, and more, it is explainable to domain experts, rather than as a score as happens with common feature selection measures. We modified an existing time intervals related patterns discovery algorithm according to the Saraswati suite, and evaluated it rigorously against the current approach on six real-life datasets. Our results show that the Saraswati-based algorithm is much faster than discovery of the entire set of frequent patterns, and the selection criteria are more effective than existing state-of-the-art feature selection methods when the discovered predictive patterns are used for classification. Additionally, the selection of the patterns is explainable in the domain expert’s terminology based on several meaningful metrics.