Multi-Time Scale Aware Host Task Preferred Learning for WEEE return prediction

Recently, with the improvement of per-capita income, the number of waste electronic and electrical equipment (WEEE) has increased significantly. The WEEE return prediction is an essential part of reverse logistics (RL) due to its helpfulness in decision-making. The traditional prediction methods usually learn from the historical data of merely a single type of WEEE. However, the prediction tasks of different types of WEEE are relevant to some extent, the lack of considering their relationships in prediction leads to sub-optimal performance. To this end, we propose a multi-task learning model, Multi-Time Scale Aware Host Task Preferred Learning model (MAHOP), to predict return volume by learning from multiple types of WEEE. The work is non-trivial due to the challenges: (1) Collaborative extraction of multi-time scale features and multi-task common features from different types of WEEE, (2) fair prediction for every type of WEEE, and (3) Proper usage of common different time scale features. To tackle these challenges, we first construct a multi-task learning framework with different towers to learn three common time-scale features from the time series data of all types of WEEE. Besides, we propose a polling host-task learning strategy and a host-preferred loss. Moreover, we design a more sharing tower to make the model not overly dependent on some specific time series features. We apply MAHOP to a WEEE recycling enterprise and conduct extensive experiments to demonstrate that MAHOP outperforms baseline models, with improved performance and acceptable hyper-parameter sensitivity. To be more specific, the average prediction error for the fridge, air conditioner, washing machine, and television is about 8% lower than that of the suboptimal model.