Dwi Ayu Lusia, Achmad Efendi, Adji Achmad Rinaldo Fernandes, Abela Chairunissa, Eka Dani Maulana, Tarisa Anggraini Purnomo, Vincent Osbert
Explainable Artificial Intelligence has gained importance alongside the increasing complexity of Artificial Intelligence models. Nonlinear Artificial Neural Network models, while powerful, are often difficult to interpret and face parameter estimation challenges when input variables are highly correlated. To address this issue, this study aims to enhance the interpretability and stability of Artificial Neural Network models by employing linear activation functions and using uncorrelated inputs. The research method involves applying Principal Component Analysis for dimensionality reduction and comparing its impact on prediction accuracy against models using original variables. Furthermore, feature importance derived from the Artificial Neural Network is compared with regression coefficients obtained from Multivariate Regression Analysis to provide a deeper understanding of variable contributions. The findings show that the proposed approach produces stable models with accurate predictions, although PCA-based dimensionality reduction increases the testing Root Mean Square Error. This research contributes to developing a more interpretable and reliable Artificial Neural Network modeling framework, while also demonstrating how feature importance analysis can be aligned with regression-based interpretations for better transparency. © 2025 IEEE.
Universitas Brawijaya, Department of Statistics, Malang, Indonesia
