Authors:
Kalleb Abreu
1
;
Julio Reis
1
;
André Santos
1
and
Giorgio Zucchi
2
Affiliations:
1
Department of Informatics, Universidade Federal de Viçosa, Minas Gerais, Brazil
;
2
R&D Department, Coopservice s.c.p.a, Reggio Emilia, Italy
Keyword(s):
Alarms, Machine Learning, Clustering, Classification, Explainable Model.
Abstract:
This paper evaluates machine learning models for the prediction of alarms using geographical clustering, exploring 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, and
clustering methodologies, namely COOP (Coopservice) and KPP (K-Means++). The XGB models emerge as
the most effective, yielding the highest AP (Average Precision) values across models based on MLP and NB.
Hyperparameter tuning for XGB models reveals default values perform well. Our model explainability analyses reveal the significant impact of geographical location (cluster) and the time interval when the predictions
are made. Challenges arise in handling dataset imbalances, impacting minority alarm class predictions. the
insights gained from this study lay the groundwork for future investigations in the field of geographical alarm
prediction. The identified challenges, such as
imbalanced datasets, offer opportunities for refining methodologies. As we move forward, a deeper exploration of one-class algorithms holds promise for addressing these
challenges and enhancing the robustness of predictive models in similar contexts.
(More)