Delivery ETA Machine Learning Model With Scenario-Aware Ranges

The patent implements a single unified machine learning model that serves multiple delivery scenarios including food delivery, grocery delivery, and package delivery. This universal model replaces the need for separate task-specific models for each delivery type, thereby reducing model complexity while maintaining the ability to handle diverse delivery scenarios. The model achieves this through a shared architecture that processes different input features and generates appropriate ETA predictions for various delivery contexts.

The patent transforms the ETA prediction from a single-point estimate to a probability distribution by changing the output parameter from a deterministic value to a stochastic representation. The model outputs a distribution of possible delivery times with associated probabilities, allowing the system to provide both a most-likely ETA and a range of possible outcomes. This parameter change enhances prediction accuracy and reliability by accounting for geographic uncertainties and external factors without requiring overly complex modeling approaches.

The patent merges multiple location-specific prediction models into a single unified model that handles all retrieval locations. This consolidation combines the training data from various locations into a comprehensive dataset, allowing the model to learn location-specific patterns while sharing common features and parameters across all locations. The unified model reduces computational resources and training time compared to training separate models for each location, while maintaining location-specific prediction accuracy through feature engineering and location-aware processing.