Continuous Regression for Graph Structured Data

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 229 papers31 conceptsfrozen 2026-10-04

Most progress in graph machine learning has been defined by discrete choices: Is this molecule toxic or safe? Is this user a bot or a human? Is this paper about physics or biology? While frameworks like the TUDataset by Morris et al. have provided a rigorous foundation for these classification tasks, a significant gap remains. Many real-world problems—from predicting the melting point of a new chemical compound to estimating the age of a patient in a medical population graph—require predicting continuous, unbounded numerical values.

The research community is currently pivoting to address this "regression gap," moving beyond categorical labels to handle the nuances of continuous variables in graph-structured data.

The Breakdown of Discrete Assumptions

The primary hurdle in moving from classification to regression is that many core graph learning mechanisms are built on the assumption of discrete "compatibility." In node classification, we often rely on homophily—the idea that connected nodes likely share the same label. Zhong et al. demonstrate this in Compatible Label Propagation, where a matrix explicitly maps how different classes attract or repel one another.

However, as Zhong et al. note, this logic fails when labels are continuous. You cannot easily build a "compatibility matrix" for an infinite range of numerical values. Similarly, collective learning frameworks like the one proposed by Hang et al. in CL+GNN use joint distributions over discrete labels to boost expressiveness. Translating these to continuous output spaces requires entirely new ways of modeling joint distributions and pseudolikelihoods that can handle the "smoothness" of numerical data rather than the "hardness" of class boundaries.

The Prompting Problem in Foundation Models

The rise of Graph Foundation Models has further highlighted this regression bottleneck. Models like One for All by Liu et al. and PRODIGY by Huang et al. use natural language and "graph prompting" to enable zero-shot learning. While it is intuitive to prompt a model with "Is this node a 'researcher' or 'student'?", it is much harder to prompt for "What is the precise kinetic rate constant of this reaction?"

As Liu et al. point out, current prompting architectures are fundamentally class-based. They struggle with unbounded target values because the semantic space used for "in-context learning" is designed to map graph structures to discrete tokens. This constraint limits the utility of even the most advanced models for high-precision engineering and physical science tasks.

Redefining Graph Metrics for Continuous Space

To solve regression, we first need to be able to measure it. Traditional metrics like homophily are binary: either neighbors match or they don't. Mueller et al. address this in Extended Graph Assessment Metrics by redefining homophily as a distance-based metric. Instead of checking for label equality, they calculate the normalized distance between a node’s continuous value and those of its neighbors.

This shift allows researchers to diagnose why a GNN might fail on a regression task. If the "continuous homophily" is low, it suggests the graph structure itself may not support the numerical trend the model is trying to learn. This diagnostic tool is essential for complex domains like the RelBench benchmark by Robinson et al., where GNNs applied to relational databases often lag behind on regression tasks compared to classification.

New Architectures for Numerical Precision

Finally, the field is seeing the emergence of architectures specifically "instructed" for regression. Mastropietro et al. introduced Edge-Wise Graph-Instructed Neural Networks (EWGINNs), which use dual sets of weights to rescale information flow. This breaks the structural symmetry that often hampers standard GNNs, allowing for more stable and precise node-level regression in chaotic or random networks.

In the chemical domain, where the stakes for regression are highest, models like Rxn Hypergraph by Shmakov et al. are building "highways" for information across disconnected molecules. While these currently excel at classification, the authors identify the transition to predicting yields and rate constants as the next major hurdle for chemical machine learning.

By moving away from the "discrete-first" mindset, this research direction aims to make graph models as effective at calculating the "how much" as they are at determining the "what."

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 23 papers
  2. Version 1 · 2026-10-04 · 29 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_continuous_property_prediction_459352c4,
  title = {continuous property prediction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-property-prediction/rdr_459352c4ebb843b09f1e275893616cdc},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_continuous_regression_for_graph_structur_459352c4_v1,
  title = {Continuous Regression for Graph Structured Data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-regression-for-graph-structured-data/rdr_459352c4ebb843b09f1e275893616cdc/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}