Portal de Periódicos da CAPES

Interpretable Spatial-Temporal Attention Graph Convolution Network for Service Part Hierarchical Demand Forecast

2019; Springer Science+Business Media; Linguagem: Inglês

10.1007/978-3-030-32236-6_52

1611-3349

Wenli Ouyang, Yahong Zhang, Mingda Zhu, Xiuling Zhang, Hongye Chen, Yinghao Ren, Wei Fan,

Time Series Analysis and Forecasting

Accurate service part demand forecast plays a key role in service supply chain management. It enables better decision making in the planning of service part procurement and distribution. To achieve high responsiveness, the service supply chain network exhibits a hierarchical structure: forward stocking locations (FSL) close to the end customer, distribution centers (DC) in the middle and center hub (CH) at the top. Hierarchical forecasts require not only good prediction accuracy at each level of the service supply chain network, but also the consistency between different levels. The accuracy and consistency of hierarchical forecasts are important to be interpretable to the decision-makers (DM). Moreover, service part demand data is the spatial-temporal time series that the observations made at neighboring regions and adjacent timestamps are not independent but dynamically correlated with each other. Recent advances in deep learning enable promising results in modeling the complex spatial-temporal relationship. Researchers use convolutional neural networks (CNN) to model spatial correlations and recurrent neural networks (RNN) to model temporal correlations. However, these deep learning models are non-transparent to the DMs who broadly require justifications in the decision-making processes. Here an interpretable solution is in the urgent demand. In this paper, we present an interpretable general framework STAH (Spatial-Temporal Attention Graph Convolution network for Hierarchical demand forecast). We evaluate our approach on Lenovo Group Ltd.’s service part demand data in India. Experimental results demonstrate the efficacy of our approach, showing superior accuracy while increasing model interpretability.