Improving structural learning in Bayesian networks: Stationarity analysis for algorithm choice
German Cuaya-Simbro, Manolo Tellez Meneses, Elías Ruiz · Data and Information Management · 2025
Structural learning in Bayesian networks is crucial for accurate modeling of complex systems. However, the performance of structural learning algorithms is significantly influenced by data characteristics . This study investigates the impact of data stationarity on the performance of structural learning algorithms and proposes a measure for selecting the most appropriate algorithm based on stationarity analysis. We compared the performance of various algorithms on both stationary and non-stationary datasets, using the KPSS test to assess stationarity. Our findings indicate that Max-Min Hill Climbing (MMHC) is particularly effective for stationary data, while Hill Climbing performs better for non-stationary data. These results highlight the importance of tailoring algorithm selection to data characteristics and provide practical guidelines for researchers and practitioners. Future research could explore the development of more adaptive algorithms and delve deeper into the relationship between data stationarity and algorithm performance.