The Effect of Preprocessing Techniques on Stacked Generalization and Stand-Alone Method for E-commerce Demand Prediction

The competition between sellers of e-commerce in Indonesia are getting more intense. One of the ways for the sellers to stay competing in the industry is by managing their product inventory through demand prediction. A lot of research related to demand prediction had been conducted. On those research, preprocessing techniques were always implemented [16] [26] [33], but there hasn’t been a statement or a main guide related to the effect of preprocessing techniques on ecommerce demand prediction, like the research that was conducted on 2022 for steel industry. In this paper, multiple scenarios will be conducted to see the effects of preprocessing techniques on stacked generalization and stand-alone method using MLP (Multi Layer Perceptron) and XGBoost (Xtreme Gradient Boosting). It is found that data scaling can increase the performance of stacked generalization with MLP on level 1. The training time will also be faster if there is a significant improvement from highly skewed to normal distribution data. With feature interaction constraint, the training time of stacked generalization and stand-alone method can become faster with the same or better prediction performance, if the constraint is in accordance with the problem. Stacked generalization, especially with XGBoost on level 1, can result in a better prediction than stand-alone MLP with hyperparameter tuning and data scaling. If the data doesn’t have a high skewness, stacked generalization method can minimalize the error made by unsuitable parameter in stand-alone XGBoost.