Improving Fish Weight Estimation with Quantile and Box-Cox Transforms: Comparative Machine Learning Models

Authors : Hatice Esen, Havvanur Taşdelen, Sefa Kucuk, Işıl Karabey Aksakallı

Pages : 581-597

View : 68 | Download : 102

Publication Date : 2025-09-30

Article Type : Research Paper

Abstract :Fish weight estimation using machine learning ensures that fish are fed appropriately, reduces labor, prevents physical harm to the fish, and saves time. In this study, Quantile and Box-Cox transformations are applied to improve the accuracy of fish weight predictions. These transformations correct the asymmetric distribution of the data and enable machine learning algorithms to generalize more effectively and produce more accurate results. CatBoost, Random Forest, Polynomial Regression, and Support Vector Regression methods were evaluated for fish weight estimation both before and after applying the transformations. The experimental results show that both the Quantile and Box-Cox transformations effectively reduce model error rates, particularly by normalizing the dataset distribution. Notably, models without transformation exhibit significant improvements in error rates after transformation is applied. The lowest Mean Absolute Error (MAE) without transformation was obtained using the CatBoost model, yielding a value of 14.002. After applying the Quantile transformation, the MAE decreased to 0.0171, while the Box-Cox transformation resulted in an MAE of 0.3302. Although both transformations contribute to error reduction, the Quantile transformation has a more substantial impact on fish weight estimation. These findings underscore the importance of data transformations in the preprocessing stage and highlight that transformation techniques are as crucial as selecting the appropriate machine learning model.
Keywords : Machine learning, Quantile transformation, Box-cox transformation, Fish weight estimation

ORIGINAL ARTICLE URL