Graph Classification Techniques and Applications

Graph classification seeks to assign labels to entire graphs by extracting and learning from their structural patterns. Early approaches relied on graph kernels, which measure similarity by counting common substructures such as walks, paths or subtrees. Subgraph mining techniques enumerate frequent patterns and deploy them as features, while boosting algorithms integrate predictive subgraphs iteratively to focus on the most informative structures. More recently, graph neural networks have emerged, learning hierarchical representations by propagating and aggregating node features across edges. Hybrid methods combine kernel-based descriptors with neural encodings to capture both global topology and local motifs. Graph classification has become indispensable in cheminformatics for predicting molecular activity, in bioinformatics for protein interface recognition, and in social sciences for community detection and user behaviour profiling. In cybersecurity, graph classification underpins malware detection by treating execution traces as graphs, and in finance it supports fraud detection through transaction network analysis. The field continues to evolve towards more scalable, data-efficient and interpretable models, addressing large-scale graphs, dynamic networks and multi-relational structures.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

One foundational advance revisited mathematical-programming-based boosting, in which a branch-and-bound search identifies candidate subgraphs and a boosting framework selects the most discriminative structures with minimal iterations. This approach outperformed simpler frequent-pattern methods by pruning uninformative regions of the search space, thus reducing computational overhead without sacrificing accuracy. Another line of work addressed scenarios where each object comprises multiple graph instances and may bear multiple labels. By treating each sample as a bag of graphs, an entropy-driven selection mechanism prioritises informative substructures over mere frequency. The method then recasts the multi-graph multi-label problem into a multi-instance multi-label framework, enabling established learners to process graph-structured inputs effectively. Empirical studies demonstrated improved classification performance on image-derived graph data and medicinal compound analysis, highlighting the value of precise structure extraction and adaptive feature selection. These diverse contributions underscore the ongoing integration of pattern mining, statistical learning and problem-specific transformations to advance graph classification capabilities.

Graph Classification Techniques and Applications publication trend

The graph below shows the total number of articles in graph classification techniques and applications across all publications each year (not limited to Nature Index journals).

Technical terms

Graph classification: The task of predicting a single categorical label for an entire graph based on its topology and node or edge attributes.

Subgraph pattern: A connected arrangement of nodes and edges within a larger graph, used as a feature to characterise graph structure.

Boosting: An ensemble learning strategy that sequentially combines weak learners, here pattern-based classifiers, to form a stronger overall predictor.

Multi-instance multi-label learning: A paradigm in which each example is represented by a collection of instances (here graphs) and is associated with potentially multiple target labels.

References

1. gBoost: a mathematical programming approach to graph classification and regression. Machine Learning (2008).
2. Multi-Graph Multi-Label Learning Based on Entropy. Entropy (2018).