A multi-source multi-layer-based transfer learning approach for forecasting customer demands of newly launched products

Abstract

Forecasting the future demand for newly launched products has been challenging for supply chain practitioners, often due to the lack of data. However, market surveys and extracting knowledge by examining similar market products to find the behaviour of a new product can be inaccurate and lead to erroneous results, which ultimately lead to a misestimation of the overall cost of a business. Meanwhile, with the advancement of artificial intelligence (AI) approaches, such as Transfer Learning (TL), this misestimation of cost can be reduced by more accurately forecasting the demand for newly launched products by seeking knowledge from the historical data of other similar products. Consequently, this paper investigates several classical AI-based TL approaches to predict customer demand for new products and stores. Thereafter, a novel Multi-Source Multi-Layer Transfer Learning approach with a Recursive Feature Elimination (MSML-TL-RFE) strategy is proposed to exploit the knowledge extraction power of the model from multiple sources for different days-ahead-prediction, distinguishing itself from the other investigated approaches. In this paper, an abstract concept of a supply chain, with information sharing among retailers, is investigated to show that such concepts can escalate the knowledge transfer ability of a system. A hierarchical two-echelon supply chain model with different attributes is developed to validate the proposed MSML-TL-RFE approach against a few other TL-based forecasting approaches. The feature-rich datasets are then transformed in such a way that they depict a hierarchical supply chain structure, allowing for the effective application of TL for forecasting consumer demand for recently introduced products. Continuing with that idea of information sharing, finding comparable sources for a quick and effective knowledge transfer procedure is investigated, considering all the peculiarities of a certain data set. MSML-TL-REF predictions and other TL-based approaches are analysed by calculating overall supply chain costs. Based on overall supply chain costs under static and dynamic lead time settings, the effectiveness and applicability of the proposed MSML-TL-RFE against traditional forecasting approaches are demonstrated. Incorporating MSML-TL-RFE with three sources improves accuracy, defined as the reciprocal of Root Mean Square Error (RMSE), from 4.83 (no TL) to 5.67 and further increases to 5.76 with additional sources, enabling more accurate predictions and reduced supply chain costs for businesses.