Estimating package arrival time via heterogeneous hypergraph neural network

Estimated Time of Arrival (ETA) for packages plays an essential role in intelligent logistics. As a classic ETA method, Origin-Destination-based (OD-based) ETA predicts the delivery time only based on the attributes (i.e., sender address, receiver address, seller, and payment time) of packages under the condition that the delivery route is unavailable. However, existing OD-based methods only exploit attributes associated with an individual order, which fails to model the higher-order interactions within orders and attributes, and fail to sufficiently exploit the graph-structure knowledge (i.e., relation of orders and attributes) and feature-based knowledge (i.e., statistical properties) of orders simultaneously, resulting in inaccurate predictions. In this paper, we propose a novel Heterogeneous HyperGraph Neural Network (H2GNN) for estimating package arrival time. Specifically, to better capture the high-order interactions within orders and attributes, we construct an order heterogeneous hypergraph that utilizes hyperedges to represent orders and nodes to represent order attributes. Besides, we extend the hypergraph learning for large-scale e-commerce data by Hyper-GraphSAGE. Overall, H2GNN can provide informatively representations of packages while preserving both structure-based knowledge learned by hypergraph and feature-based knowledge captured by Transformer. Experimental results on large-scale Alibaba logistics data demonstrate the superior performance of H2GNN compared to the baselines.