Interval-based Interpretable Decision Tree for Time Series Classification
Malte Schmidt, Volker Lohweg · 2021
In this paper we present the first iteration of a novel time series classification algorithm which is globally and inherently interpretable.The need for model interpretability or explainability is commonly agreed upon in industry [1].Model interpretability is an important characteristic of a classifier to build trust in the decisions of the classifier and makes it possible to iteratively improve a model with domain knowledge.The proposed algorithm first performs an unsupervised clustering of random segments of random length of a time series to find the most discriminating patterns.After finding segments with discriminating patterns, a decision tree is trained using the cluster labels as features.Therefore, the decision tree is restricted to learn a mapping from discriminating clusters to given class labels.The performance of our algorithm is compared to state-of-the-art algorithms with a computational feasible subset of the University of California, Riverside, time series archive [2].The first iteration of our algorithm is computationally expensive and does not achieve state-of-the-art accuracy.We point out shortcomings of the current iteration and discuss planned improvements to our algorithm to tackle these shortcomings.We find that our algorithm creates shallow decision trees which boosts interpretability.In contrast, not all stateof-the-art approaches provide interpretable models.