Impact of variable selection and model complexity on the prediction of water quality parameters for Penaeus vannamei aquaculture in a short dataset context

Abstract

Aquaculture is expanding rapidly worldwide, increasing the demand for efficient water quality management in shrimp farming. In this study, we evaluated the impact of variable selection and model complexity on the prediction of the mean of water parameters using machine learning. Two variable selection approaches were applied: a Granger causality test to capture temporal predictability, and a backward procedure based on the Akaike Information Criterion to balance model fit and complexity. An experimental dataset of 106 observations of temperature, dissolved oxygen, salinity and pH was standardised and modelled using a linear regression and a random forest regressor. Model performance was assessed by cross-validation using mean squared error (MSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) as metrics. Our results showed a significant superiority of linear regressor over the random forest, suggesting that simpler models may be more effective with limited datasets than more complex models.