Investigating the effect of feature engineering on the performance of statistical, machine learning, and deep learning models in predicting monthly milk production.
Investigating the effect of feature engineering on the performance of statistical, machine learning, and deep learning models in predicting monthly milk production.
Seyed Alireza Razavi1 reza najafi shiraye2 Goodarz Saadati Moghadam3
1) Master s student in Information Technology – E-Commerce, Islamic Revolution Comprehensive University.
2) Master s student in Information Technology – E-Commerce, Islamic Revolution Comprehensive University.
3) Assistant Professor, Department of Information Technology Engineering – E-Commerce, Islamic Revolution Comprehensive University.
IOI Code
:
ABDE-ABKFGF
Publication :
6th International Conference on Advanced Research in Management And Humanities(huconf.com)
Abstract :
Forecasting production volume is essential for improving processes and supply chain management in manufacturing companies. If the manufactured products or downstream products have expiration dates or are perishable, forecasting production volume becomes even more important. Therefore, in this study, by selecting the milk production problem and examining feature engineering, an attempt has been made to find a method that can achieve higher accuracy in predicting production volume by investigating its effect on evaluation metrics.
In this research, three feature engineering scenarios were examined. In the first scenario, only one lag was used, while in the second scenario, the full feature space including calendar features and short-term lags was used, and in the third scenario, long-term monthly lags (12, 24, and 36 months) were used. The models employed in this study are: linear regression, SVR, random forest, Gradient Boosting, XGBoost, LightGBM, CatBoost, SARIMA, Holt-Winters, Seasonal Naive, Fuzzy TS, LSTM, and Mamba. The prediction accuracy of the models was measured using the evaluation metrics MAE, RMSE, R², MAPE, and Max Error.
In the first scenario, the LSTM model achieved the best evaluation metric performance, reaching R² = 0.584. In the second scenario, R² = 0.949 was recorded for Gradient Boosting and XGBoost, which was the best result among the three scenarios. In the third scenario, XGBoost and the random forest model also achieved the best result with R² = 0.712.
Keywords : feature engineering time series forecasting milk production comparison of evaluation models statistical models