🤖 machine learning

Intelligent Truck Matching in Full Truckload Shipments using Ping2Hex approach

This paper introduces ITM 2.0, a machine learning system that leverages Uber H3 spatial indexing and LightGBM to solve the critical challenge of matching full truckload shipments to trucks despite missing identifiers, achieving significant precision gains and doubled coverage in production environments.

Original authors: Srinivas Kumar R, Jose Mathew, Ankit Singh Chauhan, Dinesh Rajkumar, Aravind Manoj, Mohit Goel

Published 2026-05-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Srinivas Kumar R, Jose Mathew, Ankit Singh Chauhan, Dinesh Rajkumar, Aravind Manoj, Mohit Goel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a logistics manager trying to track a package. You know the package left a warehouse at 8:00 AM and is heading to a specific city. You have a map showing hundreds of trucks leaving that same warehouse around that time. The problem? The digital "name tag" (like a license plate or device ID) on the trucks is either missing, broken, or written down wrong.

In the past, if the name tag was missing, the system would just give up. It wouldn't know which truck had your package, leaving you in the dark about when it would arrive.

This paper introduces a new "smart detective" system called ITM 2.0 that solves this problem without needing a name tag. Here is how it works, using simple analogies:

1. The Problem: The "Lost Name Tag"

Think of a busy truck stop where 50 trucks leave at the same time. Usually, you'd just look at the license plate to know which one has your cargo. But if the plate is covered in mud (missing data) or the driver wrote the wrong number (corrupted data), traditional systems get confused. They can't tell which truck is which, so they stop tracking the shipment entirely.

2. The Solution: The "Hexagon Map" (Ping2Hex)

Instead of trying to read the muddy license plate, the new system looks at where the trucks are going.

The researchers used a special map tool (called Uber's H3) that divides the world into a giant honeycomb of hexagons (like a beehive).

The Old Way: Tried to match exact GPS dots. If the GPS was off by a few meters (like a bad phone signal), the match failed.
The New Way: The system ignores exact dots and looks at the hexagons. It asks: "Did this truck drive through the same honeycomb cells as the trucks that usually take this route?"

Imagine you are trying to guess which friend is walking home. You don't need to see their face; you just need to see that they are walking down the same street, turning the same corners, and passing the same landmarks as your friend usually does.

3. The "Smart Brain" (Machine Learning)

The system doesn't just look at the map; it uses a "brain" (a machine learning model called LightGBM) to make a guess.

It looks at the time (did the truck leave at the right time?).
It looks at the path (is the truck driving through the same hexagons as the historical route?).
It calculates a probability score: "There is a 90% chance this truck is the one with your package."

4. The "Safety Filter" (Post-Processing)

Sometimes, the "brain" might think two different trucks both look like good matches. To avoid mistakes, the system has a strict rulebook (a post-processing step).

If the confidence score is high, it says, "Yes, that's the truck!"
If the score is too low, it says, "I'm not sure yet, let's wait for more data."
This prevents the system from guessing wrong, which is worse than not guessing at all.

5. The Results: A Huge Win

The paper tested this new system against the old "rule-based" system (which was like a rigid checklist). The results were impressive:

North America: The new system was 26% more accurate and found matches for twice as many shipments as the old system.
Europe: It was 14% more accurate and found matches for 10 times more shipments.

Why This Matters

Because this system can figure out which truck is carrying a shipment even when the ID is missing, it means:

No more blind spots: Hundreds more shipments get tracked every day.
Better predictions: You get a more accurate "Estimated Time of Arrival" because the system is actually watching the right truck.
Fewer surprises: If a truck goes off-route, the system knows immediately and can alert you.

In short, this paper describes a system that stops relying on broken ID tags and instead uses the "shape" of the truck's journey to figure out where your cargo is, making the supply chain much more visible and reliable.

Technical Summary: Intelligent Truck Matching in Full Truckload Shipments using Ping2Hex Approach

1. Problem Definition

In full truckload (FTL) supply chain operations, real-time visibility relies on accurately matching GPS pings from Electronic Logging Devices (ELDs) to specific shipments. However, traditional matching methods fail when vehicle identifiers (such as truck IDs or device IDs) are missing, corrupted, or inconsistently entered due to device errors or manual data entry gaps.

When identifiers are unavailable, shipments lose visibility, preventing accurate Estimated Time of Arrival (ETA) predictions and proactive exception management. Existing solutions often assume identifiers are present or rely on rigid rule-based proximity thresholds (e.g., a 500-meter radius from a pickup point), which are highly sensitive to geocoding errors and cannot distinguish between multiple trucks departing from the same location but heading to different destinations.

2. Methodology: The ITM 2.0 Pipeline

The authors propose Intelligent Truck Matching (ITM) 2.0, a machine learning system that reframes truck-shipment matching as a probabilistic ranking problem. The system operates through the following pipeline:

Spatial Discretization (Ping2Hex): Instead of using raw GPS coordinates, the system maps GPS pings and historical lane trajectories into Uber's H3 hexagonal spatial index.
- Resolution Strategy: Origin and destination cities are mapped using H3 resolution 4 (approx. 26 km edge length), while individual pings are mapped to resolution 6 (approx. 3.72 km edge length). This balances city-level context with precise trajectory tracking.
- Feature Extraction: The system constructs features based on the overlap between a candidate truck's current pingset and historical lane data. Key features include the number of overlapping hexcells, the total number of pings within those overlapping cells, and temporal metrics (time elapsed since scheduled pickup).
Candidate Filtering: For a given shipment, candidate trucks are identified by carrier details and filtered based on proximity to the pickup location and adherence to appointment times.
Machine Learning Model: The system employs LightGBM (Light Gradient Boosting Machine) to predict the probability that a specific candidate truck's trajectory matches the shipment. LightGBM was selected over SVM and XGBoost due to its superior performance in handling high-dimensional hexcell feature spaces, class imbalance, and its ability to produce well-calibrated probabilities.
Post-Processing: A threshold-based algorithm ranks candidate trucks. It assigns confidence labels (HIGH, MEDIUM, LOW) based on probability scores ( $\tau_{high}$ , $\tau_{medium}$ ) and a minimum threshold ( $\tau_{min}$ ). If no candidate exceeds the minimum threshold, the system waits for more data rather than making a low-confidence assignment.

3. Key Contributions

The paper outlines three primary contributions to the field of supply chain visibility:

Problem Formulation: A formal definition of GPS-based truck-shipment matching as a probabilistic ranking problem over candidate pingsets, specifically addressing the scenario where explicit identifiers are missing.
Architecture: The development of the ITM 2.0 ML pipeline, which improves coverage and accuracy over incumbent rule-based engines by integrating spatial and temporal features.
Dataset Construction: A novel scheme for dataset construction using the ping2hex method. This approach leverages hexagonal grid-based spatial feature extraction to model route similarity, utilizing historical lane data to train the model on distinguishing diverging trajectories.

4. Experimental Results

The system was evaluated through offline model selection and production shadow testing (duplicating live requests to the new system without affecting operations).

Model Selection: In offline comparisons, LightGBM outperformed SVM and XGBoost, achieving 55% precision in North America (NA) and 32% in the European Union (EU) on the validation set.
Production Performance (Shadow Testing): Compared to the rule-based baseline (ITM 1.0), ITM 2.0 demonstrated significant gains:
- North America: Precision increased by 26 percentage points (from 29% to 55%), while coverage doubled (from 4.5% to 9.0%).
- Europe: Precision increased by 14 percentage points (from 18% to 32%), with a 10x increase in coverage (from 2.5% to 27%).
Ablation Studies:
- Removing hexcell features (relying only on temporal data) caused precision to drop drastically to 7% in NA and 6% in EU, confirming that spatial discretization is the primary driver of accuracy.
- Removing post-processing reduced precision by 22–35 percentage points, highlighting the necessity of threshold-based filtering to prevent false positives.
Robustness: The system demonstrated robustness to geocoding errors up to 1 km and successfully identified the correct truck even when multiple candidates initially shared a direction before diverging.

5. Significance and Claims

The paper claims that ITM 2.0 directly addresses a critical gap in supply chain visibility where traditional identifier-based methods fail. By treating matching as a ranking problem under uncertainty, the system enables hundreds of additional full truckload shipments to gain real-time tracking daily.

The authors assert that their work establishes hexagonal spatial discretization of GPS trajectories as a viable and effective representation for freight matching under uncertainty, particularly when dealing with noisy geospatial data. The system's ability to double coverage while significantly improving precision allows for more accurate ETA predictions and proactive exception management, directly enhancing operational capabilities for carriers and customers.

The paper notes limitations, including reduced performance on short-haul shipments (≤40 km) due to high spatial overlap and a dependency on historical lane data for new routes. However, the current deployment at Project44 successfully handles medium and long-haul shipments in North America and Europe, providing a robust solution for the full truckload sector.