Multi-scale Temporal Feature Fusion for Time-Limited Order Prediction
Published in China Conference on Wireless Sensor Networks, 2022
The time-limited order is a new type of real-time delivery service that platforms need to complete the order delivery with different time granularity (e.g. one day, two days, or three days). Predicting the number of time-limited orders plays an important role for real-time order delivery allocation and anomaly detection in logistics IoT scenarios. However, the impact between orders of different time granularity is complex and the contribution of the static order features is unknown. Previous order predicting methods are not suitable for time-limited orders because they do not fully consider the dependencies among orders with different time granularity. In this paper, we propose a spatial-temporal framework based on stacked long short-term memory networks (LSTM) and deep & cross network (DCN) to take the dependencies of multi-scale temporal features into account and fuse cross-domain static features effectively. In addition, we utilize a multi-head attention mechanism to model the heterogeneous strengths of different dependencies. We evaluate our model on a real-world dataset with about 400,000 orders from one of the largest logistics companies in China. The evaluation results show that the Mean Absolute Error and R2 score of our method achieve 9.407 and 0.948, outperforming state-of-the-art solutions. Download paper here
Recommended citation: Wang, J., Zhou, X., Liu, Y., Zhang, X., & Wang, S. (2022, November). Multi-scale Temporal Feature Fusion for Time-Limited Order Prediction. In China Conference on Wireless Sensor Networks (pp. 132-144). Singapore: Springer Nature Singapore. https://link.springer.com/chapter/10.1007/978-981-19-8350-4_11