If This Do That: Interpretable Machine Learning Models For High Stakes Decision-Making

• Autor(a) principal: Vasconcelos, Lourenço Malheiro Serpa de
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/147468
• topic: Interpretability; Machine Learning; Time Series; Interpretable Machine Learning; Rule-based Algorithms; Emergency Department Waiting Times; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

In recent years, there has been an increasing trend in the use of black-box machine learning models in high-stakes decision making throughout different domains of society. Even though these models might be very accurate, there is a need to understand the reason behind the decisions, that is, a need to interpret the models. This need can be seen in predicting the waiting time in the Emergency Department of a hospital, where we can not simply trust a model prediction to manage a real health environment without understanding its reasons, thus, making black-box models impracticable in this scenario. Although multiple studies have aimed to explain black-box models, this can give the model even more power and be misleading which is why we should avoid it in highstakes decisions and use an interpretable model instead. This thesis demonstrates that interpretable machine learning models can be as good or even better than state of the art black-box models when predicting the waiting time in the Emergency Department of a hospital. Moreover, we also propose four new rulebased methods. To achieve this, we implemented 12 machine learning models, 6 noninterpretable (black-box) methods (ARIMA, SARIMA, Prophet, LSTM, GRU and the Transformer) and 6 rule-based and interpretable methods (RuleFit, SIRUS, REN, RDLL, RDLE and RDLR), 4 of which are proposed in this work (the REN, RDLL, RDLE and RDLR). The results obtained revealed that although some black-box models can achieve very good predictions of the waiting times in the emergency department, in some scenarios, interpretable models can outperform them.