Introduction
Estimating the Time of Arrival (ETA) is a critical component of modern logistics, yet most existing research focuses on point-to-point travel. Package delivery presents a more complex "last-mile" challenge: a single courier must navigate a sequence of dozens or even hundreds of destinations. The delivery time for any given package is intrinsically linked to the route taken for all preceding stops. This problem is compounded by time-variant courier efficiency and the need to integrate diverse features ranging from package weight to individual delivery habits.
The DeepETA Framework
To address these challenges, the authors developed DeepETA, an end-to-end neural network architecture designed to capture the spatial-temporal dependencies of multi-destination delivery routes.
The model consists of three primary components:
- Latest Route Encoder: This module represents the courier's current journey. It utilizes Geohash encoding to map geographic coordinates into strings, ensuring that proximal locations share similar vector representations. These embeddings are processed by a Bidirectional LSTM (BiLSTM) to extract sequential features.
- Frequent Pattern Encoder: Recognizing that couriers are "creatures of habit," this module uses an attention mechanism to query historical data. It distinguishes between has-delivered patterns (completed routes) and to-be-delivered patterns (remaining destinations). For the latter, since the exact future sequence is unknown, the model treats them as an unordered set using a standard Deep Neural Network (DNN).
- Joint Prediction Module: This final stage integrates the route representations, historical patterns, and time-invariant features (e.g., courier ID, package info, weather) through a fully connected layer to produce the final ETA.
A key innovation is the use of a hybrid loss function combining Mean Square Error (MSE) and Mean Absolute Percentage Square Error (MAPSE), which improves robustness against outliers like failed delivery attempts.
Experimental Results
DeepETA was evaluated on a large-scale dataset from Cainiao Ltd., comprising 350,000 samples from 331 couriers in Beijing. The model demonstrated superior performance compared to traditional regression models and specialized sequential models like DeepMove.
DeepETA achieved an RMSE of 63.58 minutes and a MAPE of 20.6%, marking a 13.8% improvement over the best-performing baseline. The results indicate that as a courier progresses through their route and uncertainty regarding the remaining stops decreases, the prediction error drops significantly—reaching as low as 29 minutes for the final packages.
Conclusion
DeepETA successfully models the complex, multi-destination nature of last-mile delivery by combining spatial-temporal encoding with historical pattern matching. By treating the future route as an unordered set, the model avoids the pitfalls of predicting exact sequences while still capturing essential geographic context. Future work may involve sequence-to-sequence modeling to predict arrival times for all remaining packages simultaneously.
Summary updated