Authors:
Ilija Šimić
;
Christian Partl
and
Vedran Sabol
Affiliation:
Know-Center GmbH, Graz, Austria
Keyword(s):
Explainable AI, Interactive Systems, Deep Learning, Attribution Methods, Visualization, Time Series, Recommender.
Abstract:
The rising popularity of black-box deep learning models directly lead to an increased interest in eXplainable AI - a field concerned with methods that explain the behavior of machine learning models. However, different types of stakeholders interact with XAI, all of which have different requirements and expectations of XAI systems. Moreover, XAI methods and tools are mostly developed for image, text, and tabular data, while explainability methods and tools for time series data - which is abundant in high-stakes domains - are in comparison fairly neglected. In this paper, we first contribute with a set of XAI user requirements for the most prominent XAI stakeholders, the machine learning experts. We also contribute with a set of functional requirements, which should be fulfilled by an XAI tool to address the derived user requirements. Based on the functional requirements, we have designed and developed XAIVIER, the eXplainable AI VIsual Explorer and Recommender, a web application for
interactive XAI in time series data. XAIVIER stands out with its explainer recommender that advises users which explanation method they should use for their dataset and model, and which ones to avoid. We have evaluated XAIVIER and its explainer recommender in a usability study, and demonstrate its usage and benefits in a detailed user scenario.
(More)