A joint feature updating training method based on time-series prediction

Inhalt:
In order to improve the performance and accuracy of single prediction model in multi category products sales forecasting, a joint feature updating training method of multi category sales forecasting based on ensemble learning is proposed in this paper. This training method which using Stacking strategy combined with Catboost, Decision Tree, XGB Regression can learn more data points than the single model. The model divides data sets into a category part and a numerical part. A label encoder or other category encoders could be applied to process the category part, whereas time-series feature methods could be applied to process the numerical part. After that, an iterative predicting method is involved in making long-term predictions. The method includes a feature updater that helps generate the feature for the next predicting step. The Retail Data Analytics data set from Kaggle was used to test the efficiency of the proposed model. The results shows that the proposed model had high efficiency and predicting accuracy.