DT-MUSA: Dual Transfer Driven Multi-source Domain Adaptation for WEEE Reverse Logistics Return Prediction

Reverse logistics (RL) return prediction for Waste Electrical and Electronic Equipment (WEEE) has gained attention due to its potential to improve operational efficiency in the recycling industry. However, in data-scarce regions, commonly used deep learning models perform poorly. Existing multi-source cross-domain transfer learning models can partially overcome data scarcity by using historical data from multiple sources. However, these models aggregate multi-source domain data into a single-source domain in transfer, ignoring the differences in time series features among source domains. Additionally, the lack of historical data in the target domain makes fine-tuning the prediction model inoperative. To address these issues, we propose Dual Transfer Driven Multi-Source domain Adaptation (DT-MUSA) for WEEE RL return prediction. DT-MUSA includes a dual transfer model that combines sample transfer and model transfer and a basic prediction model MUCAN (Multi-time Scale CNN-Attention Network). It employs a multi-task learning to aggregate predictors from multiple regions and avoids negative transfer learning. The dual transfer model enables fine-tuning of the base model MUCAN by generating long-term time series data through sample transfer. We applied DT-MUSA to real cases of an RL recycling company and conducted extensive experiments. The results show that DT-MUSA outperforms baseline prediction models significantly.