계절조정을 활용한 전이학습 기반의 자동차 예비 부품 장기수요예측

Abstract

In this paper, we propose a TSLF methodology, which is a Transfer learning with Seasonal adjustment for Long- term Forecasting. The lack of learning data has been a chronic problem in the field of long-term demand forecasting for automotive spare parts which tends to lead to the over-fitting. To solve this problem, we used transfer learning. Transfer learning is actively used in various industries as a technique for reusing pre-trained models to improve the performance of tasks with small data, but there are no research cases applying it. The main idea is to utilize trends that are highly related to other parts as the source domain, and to migrate the pre-trained network while retaining the feature extraction layer. Experiments show that this method mitigates over-fitting problem and reduces error by 14.85% on MAE and 11.3% on RMSE than the traditional method in the small data set.