Erol B.Inkaya T.2024-03-132024-03-1320239798400709098https://doi.org/10.1145/3631908.3631917https://hdl.handle.net/20.500.12662/28957th International Conference on Algorithms, Computing and Systems, ICACS 2023 -- 19 October 2023 through 21 October 2023 -- -- 197097Sales forecasting is one of the most important tasks in supply chain management. With effective and accurate sales forecasts, the cost of managing the supply chain can be reduced while increasing customer satisfaction. Therefore, forecasting future sales is of great importance for the efficiency of supply chains. Today, the use of deep learning approaches in sales forecasting has been prominent due to their capability of handling complex and non-linear relations in the data. However, deep learning models need to be trained with sufficient data. As a solution to the case of insufficient data, the concept of transfer learning is used, which enables the transfer of information obtained while solving a problem to another problem. In this study, we present two ensemble deep transfer learning approaches for sales forecasting, namely bagged and stacked deep transfer learning approaches. In the proposed approaches, long short-term memory (LSTM) network has been used so that long-term dependencies can be captured. Then, LSTM based transfer learning (LSTM-TL) has been performed. Finally, bagging and stacking ensemble strategies are applied to increase the performances of the LSTM-TL models. The performances of the two proposed approaches were evaluated using benchmark sales datasets. In the experimental studies, the proposed approaches were compared with base LSTM and LSTM-TL models in terms of prediction accuracy and training times. The results show that the use of ensemble learning improved the prediction accuracy of the deep learning models. In particular, the stacked deep transfer learning approach yielded better accuracy values compared to other competing approaches including the bagged deep transfer learning approach. The proposed approaches can be used in predicting other time series problems as well. © 2023 Copyright held by the owner/author(s). Publication rights licensed to ACM.eninfo:eu-repo/semantics/closedAccessConference Object10.1145/3631908.36319172-s2.0-8518553127866N/A60