Editorial: Human-Interpretable Machine Learning

Gabriele Tolomei, Fabio Pinelli, Fabrizio Silvestri

Abstract: This Research Topic encouraged submissions that broadly address the challenge of making machine learning (ML) models more transparent and intelligible to humans. Indeed, this follows the spirit of the explainable AI (XAI) initiative Samek et al. (2019), which promotes efforts to improve the humaninterpretability of ML systems, especially those supporting decisions in social domains like finance Bracke et al. (2019) and healthcare Ahmad et al. (2018).In this Research Topic, Guidotti and D'Onofrio (2021) propose MAPIC, a novel and efficient method to train time-series classification models that are natively interpretable by design based on matrix profiles (i.e., roughly, distances between all the time-series subsequences and their nearest neighbors). Time-series classification is a pervasive and transversal problem in various domains, ranging from disease diagnosis to anomaly detection in finance. Inspired by previous work on time-series classifiers based on shapelets (Ye and Keogh, 2009;Trasarti et al., 2011), MAPIC operates as follows. First, to find the best shapelets, MAPIC exploits the matrix profiles extracted from the time-series of the training set instead of using a brute force approach (Ye and Keogh, 2009) or an optimized search (Grabocka et al., 2014). Hence, MAPIC retrieves motifs and discords from the matrix profiles of each time-series and adopts them as candidate shapelets. Second, differently from traditional approaches that learn machine learning models for ...