Delivery Time Inference Based on Couriers' Trajectories

Efficient and reliable logistics are essential for e‑commerce, and the manual confirmation of delivery ("proof of delivery") by couriers creates a heavy burden. This work introduces the paper "Doing in One Go: Delivery Time Inference Based on Couriers' Trajectories" presented at KDD2020, which describes how JD's spatiotemporal data engine (JUST) can reduce that burden.

Problem Background – Each parcel generates an order (waybill) that records customer, package, and delivery status information, including the time the package is handed to the customer (delivery time). Accurate delivery‑time data supports order lifecycle management, receipt prediction, behavior analysis, and insurance planning.

Challenges – (1) The geocoded address coordinates are not always the true delivery location due to geocoding errors and varied customer receipt methods, causing a spatial offset of several meters. (2) Couriers may pause for phone calls, traffic lights, or other reasons, producing stay points that are not actual delivery locations.

Proposed Solution (DTInf) – The DTInf pipeline consists of three modules:

1) Data Pre‑processing : Noise filtering of raw GPS traces, detection of stay points, and segmentation of waybills and stay points according to identified delivery trips.

2) Delivery‑Location Correction : Mining historical trips to build a mapping from geocoded coordinates to true delivery locations, using the centroid of stay‑point clusters and spatial clustering to reduce redundancy.

3) Delivery‑Event Matching : Grouping waybills by corrected delivery location into delivery events, then applying a pre‑trained stay‑point selection model (a multilayer perceptron) that scores candidate stay points based on delivery‑location features, event features, stay‑point features, and spatial distance, selecting the most probable stay point as the inferred delivery time.

Experiment – The method was evaluated on 5 couriers' GPS data (593 M points) and 274 k waybills from JD Logistics (Beijing) over 15 months. Using the first 80 % of trips for training, 10 % for validation, and 10 % for testing, DTInf achieved a mean absolute error (MAE) of 401 seconds, improving over baseline methods by 31.8 %.

System Deployment – The inference system is integrated into the JUST spatiotemporal engine and can infer delivery times for all successful deliveries with only a few failed‑delivery confirmations from couriers.