Evaluating deep models for absenteeism prediction of public security agents

摘要

Absenteeism is a complex phenomenon characterized by the physical absence of the individual, usually at his workplace. Such absences generally lead to innumerable personal, social, and economic losses, particularly in public security institutions, where incidence is higher than the one verified in other occupational categories. Identifying preponderant absenteeism factors and allowing preventive actions to be carried out effectively may be beneficial to these institutions and their agents. Such knowledge could be acquired hypothetically by exploiting large human resources data sets. In this paper, we investigate the potential of machine learning classifiers to identify security workers prone to long-term absenteeism. Such predictors shall make decisions based on the professional history of each agent, which is extracted from databases of public security institutions. In our study, we performed experiments on a database comprised of 6 years of professional data from workers of the Military Police of Alagoas, Brazil. We evaluated deep models, including variations of Multilayer Perceptrons (MLP), Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM), and compared with baseline Support-Vector Machines (SVM) classifiers. We show results revealing that the best architectures achieve up to 78% of accuracy. Also, experiments indicated that the use of data accumulated over several years improves the accuracy of the prediction of absenteeism. Finally, we conclude that such results encourage the usage of deep learning techniques to predict absenteeism and support the implementation of effective prevention measures in these institutions. (C) 2020 Elsevier B.V. All rights reserved.