Hyper-parameter selection in machine learning methods

Hyper-parameters are critical for machine learning algorithms as they have a significant impact on the performance of machine learning models. Failure to choose appropriate hyper-parameters can cause machine learning models to underperform. Within the scope of this thesis, the effects of different hyper-parameter optimization (HPO, Hyper- parameter Optimization) methods on the performance of regression models are given through comparative analysis. In order to evaluate the proposed methods within the scope of the thesis, the energy efficiency problem in buildings is discussed. In the study, a dataset containing 768 lines and 8 features, which is available in the literature in this field, was used. The work began by training regression models using default parameter values. Then, among the performances of these models, the models that gave the best results were determined according to the mean squared error (MSE) value. The identified models were subjected to hyper-parameter optimization for further improvement. As a result, category boosting (CatBoost, Category Boosting) and Bayesian Optimization (Random Forest) algorithms achieved 0.997 R-Square and 0.23 MSE scores, the highest in this study, with the optimization of hyper-parameter values. was achieved as a result of success. This study addresses the energy efficiency problem in buildings to demonstrate the impact of hyper- parameter optimization methods on machine learning models. In this context, the critical role of hyper-parameter optimization in increasing the success of machine learning models is emphasized.