Explainable machine learning for effective alarm prediction

• Ano de defesa: 2023
• Autor(a) principal: Abreu, Kalleb Morais de Moura
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Viçosa; Ciência da Computação
• Palavras-chave em Português: Aprendizado do computador; Alarmes; Análise por agrupamento; Modelos e construção de modelos; Ciência da Computação
• Link de acesso: https://locus.ufv.br/handle/123456789/32764; https://doi.org/10.47328/ufvbbt.2024.242

This dissertation evaluates twelve machine learning models for the prediction of alarms using geographical clustering, leveraging data from an Italian company. The models encompass a spectrum of algorithms, including Naive Bayes (NB), XGBoost (XGB), and Multilayer Perceptron (MLP), coupled with encoding techniques such as Label/Ordinal Encoding (LOE) and Label/Ordinal/One-Hot Encoding (L2OE), and clustering method- ologies, namely Coopservice-2022 (COOP) and K-Means++ (KPP). XGB emerges as the most effective, yielding the highest AUC values across models. Adjustments in encoding methods show significant improvements for NB and MLP, with a marginal impact for XGB. Hyperparameter tuning for XGB models reveals default values outper- form varied configurations. The SHAP value analyses emphasize the significant impact of a specific cluster and hour of the day. Transfer learning experiments confirm the model’s adaptability across Italian provinces, with continuous monitoring essential due to sensitivity to cluster labels. Challenges arise in handling dataset imbalances, impacting minority alarm class predictions. This work sets a foundation for further research on specific approaches for dealing with imbalanced datasets and one-class algorithms. The study advocates for ongoing validation across diverse provinces, emphasizing nuanced analyses and improvements in model robustness. Keywords: Alarms; Machine learning; Clustering; Explainable models; Transfer learning.