Improving Origin Snap Accuracy in Amap Navigation Using Machine Learning

Overview: Amap, a leading Chinese navigation solution provider, relies heavily on accurate origin snapping ("起点抓路") to deliver reliable route planning. This article introduces how Amap enhanced the accuracy of origin snapping by incorporating machine‑learning models.

What is Origin Snapping? Origin snapping binds the user’s reported start location to the actual road segment they are on. Users can select the start point in three ways: (1) manual point selection, (2) POI (Point of Interest) selection, and (3) automatic positioning via GPS, cellular or Wi‑Fi. Manual and POI selections are generally accurate, while automatic positioning can drift by several meters to hundreds of meters, leading to mismatches between the reported location and the true road.

Why Introduce Machine Learning? Previously, origin snapping relied on handcrafted rules that weighted distance, angle, speed, etc. As request volume grew, rule‑based systems showed three main drawbacks: (1) limited coverage of edge cases, (2) slow incorporation of new features, and (3) high dependence on expert knowledge. In the era of big data and AI, automating this process with data‑driven models became essential.

Rule thresholds and weights are hard‑coded and cannot adapt quickly.

Feature updates require manual intervention.

Personnel turnover hampers consistent rule tuning.

Machine learning offers stronger expressive power to capture complex relationships between features and the correct road segment, leveraging Amap’s abundant planning and real‑travel data.

How the Machine‑Learning Solution Was Built

The development followed four key steps:

Problem Definition: The task was framed as a supervised ranking problem. Given a user’s estimated location A, a set of candidate road segments B is retrieved, and the goal is to rank them so that the true road C appears at the top.

Data Acquisition & Feature Engineering: Two data types were collected:

Ground‑truth data: actual road information derived from users’ real‑travel trajectories that intersect the planning request.

Feature data: three categories of features – location‑based, road‑based, and interaction features between location and road. Feature cleaning included deduplication, outlier handling, and value correction.

Model Selection: Since the problem is a ranking task, a list‑wise learning‑to‑rank approach was chosen, using tree‑based models. The evaluation metric was NDCG (Normalized Discounted Cumulative Gain), which measures ranking quality.

Model Training & Evaluation: A labeled dataset was split into training and test sets. Both the legacy rule‑based system and the ML model were run on the test set and compared against ground truth. The ML model reduced the error rate by 10% and improved accuracy by 40% on the discrepant cases.

Conclusion

The integration of big data and machine learning into Amap’s origin snapping pipeline significantly boosted accuracy and streamlined the workflow. Future work will focus on further refining the model, exploring new profit points, and continuously optimizing the learning‑to‑rank system from both data and algorithm perspectives.