Improving Food Delivery ETA Prediction with GBDT Feature Construction

ETA (Estimated Time of Arrival) is a critical variable in food‑delivery services, directly affecting user experience, order retention, and rider satisfaction.

The delivery business architecture relies on ETA as a key parameter for real‑time order dispatch, resource allocation, and performance assessment.

Compared with ride‑hailing, food‑delivery ETA is more complex because it involves three parties (rider, merchant, user) and four operational stages (order receipt, restaurant arrival, food pickup, delivery). This adds many uncertain factors such as rider behavior, restaurant speed, capacity planning, and weather.

Typical regression models for ETA include Gradient Boosted Decision Trees (GBDT), Random Forest (RF) and linear regression. While linear models are scalable, they require extensive feature engineering.

Feature engineering is crucial. GBDT can be used to generate high‑dimensional binary features by recording the leaf node index of each tree for a sample. Each leaf index is transformed into a one‑hot vector, and the concatenated vectors become new features.

The workflow is: train a GBDT on offline features, extract leaf indices for each sample, apply OneHotEncoder to obtain binary features, and combine them with the original offline and online features (merchant profile, area capacity, real‑time order flow, weather, etc.).

Hyper‑parameters such as number of trees, tree depth (typically 4‑6), and learning rate are tuned via GridSearch + Cross‑Validation to balance feature richness and over‑fitting.

Evaluation metrics go beyond standard MAE: N‑minute accuracy (probability that absolute error ≤ N minutes) and N‑minute business punctuality (probability that delivery is within N minutes of the estimate, considering early deliveries as punctual).

Experiments on Meituan’s delivery data show that adding GBDT‑generated features reduces MAE by 3.4% overall, improves N‑minute accuracy by 2.2 points, and yields larger gains on high‑value orders (MAE ↓ 2.56%, N‑minute punctuality ↑ 3.46%). Similar improvements are observed during lunch‑peak periods.

In summary, enriching ETA models with GBDT‑derived one‑hot features significantly boosts prediction accuracy. Future work includes incorporating temporal relationships, deep learning, and reinforcement learning to further raise the performance ceiling.