A Machine Learning Approach to Identifying Key Predictors of Peruvian School Principals' Job Satisfaction

Luis Alberto Holgado-Apaza^1*Dany Dorian Isuiza-Perez¹NELLY JACQUELINE ULLOA GALLARDO¹Yban Vilchez-Navarro¹Ruth Nataly Aragon-Navarrete²Wilian Quispe Layme³Danger David Castellon-Apaza¹Remo Choquejahua-Acero⁴Jaime Cesar Prieto-Luna¹

• ¹Departamento Académico de Ingeniería de Sistemas e Informática, Universidad Nacional Amazónica de Madre de Dios, Puerto Maldonado 17001, Peru, Puerto Maldonado, Peru
• ²Departamento Académico de Ecoturismo, Universidad Nacional Amazónica de Madre de Dios, Puerto Maldonado 17001, Peru, Puerto Maldonado, Peru
• ³Departamento Académico de Educación, Universidad Nacional Amazónica de Madre de Dios, Puerto Maldonado 17001, Peru, Puerto Maldonado, Peru
• ⁴Departamento Académico de Ingeniería Estadística e Informática, Universidad Nacional del Altiplano-Puno, Puno 21001, Peru, Puno, Peru

School principals encounter contemporary demands that impact their job satisfaction and leadership effectiveness. Despite the significance of this issue, there is limited research on satisfaction predictors for these professionals, particularly using machine learning approaches.This study identified key predictors of job satisfaction among Peruvian school principals by applying an ensemble of feature selection methods and evaluating five machine learning algorithms (Random Forest, Decision Trees-CART, Histogram-Based Gradient Boosting, XGBoost, and LightGBM) with data from the 2018 National Survey of Directors. The principal variables identified included satisfaction with salary, geographic location of the educational institution, relationships with students and teachers, workplace climate, student learning achievements, and job benefits. Economic factors proved important, such as gross and net income, and the minimum monthly amount required to meet household needs. Timerelated aspects also exerted influence, including hours dedicated to training, time spent on administrative and/or teaching duties outside working hours, travel time to and from the Local Educational Management Unit (UGEL), duration of stays at the UGEL, and commuting time from principal residence to the educational institution. The Histogram-Based Gradient Boosting algorithm, optimized with Bayesian techniques and trained with data balanced through Random Oversampling, achieved a balanced accuracy of 0.63 on a test set with real-world class distribution. 1 Holgado-Apaza et al.When using Generative Adversarial Networks to balance only the training set, better results were obtained in recall (0.74), precision (0.72), and F1 score (0.70). SHAP analysis revealed that economic factors primarily influenced dissatisfied principals, while interpersonal factors were more important for highly satisfied principals, suggesting a hierarchical pattern of needs. The findings could inform strategies to enhance principals' job satisfaction and strengthen data-driven educational policies.