Pengembangan Model Regresi dan Machine Learning untuk Meningkatkan Akurasi Prediksi pada Data Multivariat

Abstract

ABSTRACT

Purpose: This article develops a comparative framework for linear regression, regularized regression, and machine-learning models to improve prediction in multivariate datasets while maintaining transparent assessment of uncertainty.

Design/Methodology: A methodological and analytical literature review examines ordinary least squares, ridge, lasso, elastic net, random forests, gradient boosting, and support vector regression. A reproducible experimental protocol is proposed using preprocessing pipelines, nested cross-validation, and MAE, RMSE, and R-squared; no unperformed empirical experiment is reported.

Findings: Prediction quality depends on variable relationships, sample size, collinearity, nonlinear patterns, and validation design. Regularization can stabilize estimates when predictors are correlated, whereas nonlinear learners may capture interactions but require tuning and controls against overfitting and data leakage.

Practical Implications: Researchers can apply the framework to select appropriate baselines, evaluate models on unseen data, assess subgroup performance, and report sensitivity before deployment.

Originality/Value: The contribution is an integrated evaluation framework linking statistical interpretability, predictive accuracy, feature stability, reproducibility, and domain-aware decisions.