Optimizing Logistics Inbound Processes with Machine Learning Models

Coupang used data to predict the number of trucks needed for inbound shipments to fulfillment centers, enabling more efficient allocation of limited dock slots.

From over 2 years of collected inbound request data, the team built approximately 800,000 training samples, and through domain interviews and EDA, extracted features effective for prediction.

Given the large number of categorical features and the importance of training/tuning speed, the team chose LightGBM for the model. The trees used leaf-wise growth to speed up training, and categorical variables were handled without separate one-hot encoding.

Hyperparameter search was automated using Bayesian Optimization.

• Initially, models were trained on random samples from the search space
• Based on the results, the objective function was estimated using a Gaussian Process
• Promising next candidates were repeatedly selected to find the optimal combination

The trained model was connected to a SageMaker endpoint and the inbound reservation system. When a vendor submits a reservation, the system immediately returns the appropriate number of trucks, designed to be reflected directly in the reservation decision.

In prediction, balancing underestimation and overestimation was important. The final model achieved 2.53% underestimation and 5.04% overestimation, a significant improvement over the previous 8.71% / 44.45%.

After deployment, cases of changing inbound dates due to slot shortages decreased by 67.9%. This allowed Coupang to reduce unnecessary resource waste, while vendors could inbound more reliably on their desired schedules.

Going forward, the team plans to continuously discover derived features and add data in line with expanding product categories and changes in work methods, to keep improving prediction performance.