Purchase Forecast of the Promotion Day Based on Model Fusion and Migration Learning

Abstract

For the violent data fluctuation of the promotion day data and the characteristics of limited data size, combined with the idea of migration learning, two modeling ideas of parallel model and serial model are proposed. The algorithm avoids the difference in data distribution between promotion day and non-promotion day, and makes full use of non-promotion day data to improve the accuracy of model prediction. Experiments conducted by real promotion day data show that migration learning can effectively reduce prediction errors.