Comparing Newton Raphson and Stochastic Gradient Descent Methods for Traffic Accident in Malang

Abstract

This study discusses a comparison between two optimization methods, Newton–Raphson and Stochastic Gradient Descent (SGD), in binary logistic regression modeling to analyze the severity of traffic accidents in Malang Regency. Parameter estimation was carried out using both methods to assess their effectiveness in achieving convergence and producing a well-fitted model. The results show that the Newton–Raphson method failed to achieve convergence despite its fast iteration speed, while the SGD method successfully converged, although it required a large number of iterations. Model evaluation was conducted by examining model fit through log-likelihood values and the Akaike Information Criterion (AIC). The results indicate that the SGD method produced a better-fitting model compared to Newton–Raphson. Additionally, the regression models from each method identified different predictor variables as significant, suggesting that the choice of optimization approach can influence analytical outcomes. These findings highlight the importance of selecting an appropriate optimization method in logistic regression analysis, particularly for complex and imbalanced accident data.