Graph Neural Networks vs. Gradient Boosting: Predicting Arsenic in US Water

In July 2026, a group of researchers published a paper on arXiv that predicts arsenic concentrations in US groundwater using graph neural networks (GNNs). To do so, the authors built a single dataset from more than 74,000 water samples.

Why Arsenic in Water Is a Problem

Arsenic contamination of groundwater remains a long-standing health threat in the US, especially for households that drink water from private wells. These wells fall outside centralized water quality monitoring, so families often don't know about the danger. An accurate, spatially resolved estimate of concentrations is needed to find high-risk zones and direct remediation efforts there. The problem is that, until now, there have been no general-purpose models describing how concentrations shift smoothly from region to region.

How the Dataset Is Built

The authors merged more than 74,000 arsenic measurements from three open sources and linked them by geographic coordinates. They framed the task as regression: predicting a continuous concentration value at a specific location, rather than a coarse "dangerous/safe" label.

• More than 74,000 arsenic samples from across the US
• Sources: Water Quality Portal (WQP), Mineral Resources Data System (MRDS) and the gNATSGO soil database
• Measurement points were stitched together using k-nearest neighbors (k-NN) and GIS tools
• The task type is concentration regression, not binary classification

According to data from the Water Quality Portal and related databases, a dataset of this size allows models to train on real spatial variation rather than on isolated local samples.

Did Neural Networks Beat Boosting?

Graph neural networks matched or outperformed gradient boosting, while boosting itself remained the strongest baseline approach for tabular data. The authors compared several model families: tree ensembles (gradient-boosted trees), multilayer perceptrons (MLPs) and spatially aware GNNs. The key advantage of the graph approach is that it directly accounts for spatial dependence between neighboring points, something tabular models by their nature ignore.

"Although gradient boosting is still considered state-of-the-art for

tabular data, graph neural networks can additionally account for spatial dependence and match or outperform boosting," says the study's abstract on arXiv.

What This Means

Spatially informed learning makes environmental forecasts more accurate and provides a foundation for more reliable groundwater contamination risk maps. For millions of households with private wells, this could eventually mean earlier warnings about hazardous areas and targeted remediation where arsenic genuinely exceeds the limit.

FAQ

What is a graph neural network in this task?

It is a model that represents measurement points as graph nodes and connects geographically neighboring samples with edges. This lets the GNN account for spatial dependence, meaning how the arsenic concentration at one point relates to the points around it.

How many samples did the researchers use?

More than 74,000 arsenic measurements collected from three open databases: Water Quality Portal, Mineral Resources Data System and the gNATSGO soil database, linked by coordinates via k-NN and GIS.

Will GNNs replace gradient boosting?

No. According to the paper's conclusions, gradient boosting remains state-of-the-art for tabular data, while graph neural networks complement it by accounting for spatial structure, in places matching or outperforming it.