Human-Interpretable Explanations for Black-Box Machine Learning Models: An Application to Fraud Detection

• Autor(a) principal: Balayan, Vladimir
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/130774
• topic: Machine Learning; Explainable AI; Domain Knowledge Explanations; Self- Explainable Methods; Human-AI Cooperative systems; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

Machine Learning (ML) has been increasingly used to aid humans making high-stakes decisions in a wide range of areas, from public policy to criminal justice, education, healthcare, or financial services. However, it is very hard for humans to grasp the rationale behind every ML model’s prediction, hindering trust in the system. The field of Explainable Artificial Intelligence (XAI) emerged to tackle this problem, aiming to research and develop methods to make those “black-boxes” more interpretable, but there is still no major breakthrough. Additionally, the most popular explanation methods — LIME and SHAP — produce very low-level feature attribution explanations, being of limited usefulness to personas without any ML knowledge. This work was developed at Feedzai, a fintech company that uses ML to prevent financial crime. One of the main Feedzai products is a case management application used by fraud analysts to review suspicious financial transactions flagged by the ML models. Fraud analysts are domain experts trained to look for suspicious evidence in transactions but they do not have ML knowledge, and consequently, current XAI methods do not suit their information needs. To address this, we present JOEL, a neural network-based framework to jointly learn a decision-making task and associated domain knowledge explanations. JOEL is tailored to human-in-the-loop domain experts that lack deep technical ML knowledge, providing high-level insights about the model’s predictions that very much resemble the experts’ own reasoning. Moreover, by collecting the domain feedback from a pool of certified experts (human teaching), we promote seamless and better quality explanations. Lastly, we resort to semantic mappings between legacy expert systems and domain taxonomies to automatically annotate a bootstrap training set, overcoming the absence of concept-based human annotations. We validate JOEL empirically on a real-world fraud detection dataset, at Feedzai. We show that JOEL can generalize the explanations from the bootstrap dataset. Furthermore, obtained results indicate that human teaching is able to further improve the explanations prediction quality.