Conceptual challenges for interpretable machine learning

Journal article

Dissertation presented as the partial requirement for obtaining a Master's degree in Data Science and Advanced Analytics, specialization in Data ScienceThe adoption of Artificial Intelligence in the financial sector has brought numerous quality of life improvements on internal procedures and on decision making processes. This adoption means that AI systems actively have a say on decisions that can affect people’s lives. Meanwhile, most AI systems are opaque, casting doubts and uncertainty regarding the usage of these systems to make impactful decisions. This thesis focuses on using Explainable Artificial Intelligence (XAI) techniques and algorithms to provide explanations to an AI system’s decision and the steps it took to get there. Through a detailed literature review where the current state of XAI in the financial sector is revised and a practical use case where the LIME and SHAP algorithms are tested against an AI system developed to predict a person’s credit risk, this study tests the applicability of XAI techniques for explaining an AI system. The results suggest that the implementation of XAI techniques can provide a satisfactory degree of explainability to a model, demystifying its decision making processes

Authors: Watson, DS

• Publication status: Published
• Peer review status: Peer reviewed
• Publisher: Springer
• Journal: Synthese
• Volume: 200
• Issue: 2
• Pages: 65
• Article number: 65
• Publication date: 2022-03-01
• DOI: 10.1007/s11229-022-03485-5
• EISSN: 1573-0964
• ISSN: 0039-7857
• Language: English
• Keywords: Counterexample; Philosophy of science; Epistemology; Philosophy; Counterintuitive; Counterfactual thinking; Sociology; Ambiguity; Philosophy of language; Mathematics; Artificial intelligence; Skepticism; Computer science; Metaphysics; Machine learning; Counterfactual conditional; Point (geometry); Process (computing)