Graph Learning and Topology Inference Techniques

Graph learning and topology inference techniques aim to reconstruct network structures from observational data by treating measurements as signals defined on unknown graphs. These approaches draw on statistical signal processing, optimisation theory and machine learning to estimate adjacency matrices or graph shift operators that best explain observed patterns. Key paradigms include smoothness-based methods, which posit that signals vary gradually across connected nodes; stationarity-based frameworks, which exploit commutation between the graph shift operator and signal covariance; and sparsity-driven models, which impose that each node interacts with only a few neighbours. Structural equation and Gaussian graphical models further integrate causal or probabilistic assumptions, leading to convex formulations with sparsity regularisation. Recent advances focus on dynamic networks, developing online and time-varying algorithms that update graph estimates as new data arrive, often with provable convergence and dynamic regret bounds. Applications span neuroscience, social network analysis, sensor arrays and transportation systems. Foundational techniques recover eigenstructure from diffused signals to infer latent topologies, while modern extensions incorporate kernel methods and hidden-node modelling to handle nonlinearity and partial observations, making graph learning a versatile toolkit for uncovering complex relational patterns in high-dimensional data.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Research on dynamic graph learning has explored algorithmic frameworks that adapt to non-stationary environments. One approach casts time-varying network estimation as a composite optimisation problem, where iterative updates of empirical covariance matrices drive online recovery of Gaussian graphical models, structural equation models and smoothness-based graphs, with low-dimensional vectorisation schemes enabling efficient gradient computations. Another strand introduces kernel-based online topology estimation for nonlinear graph-connected time series, leveraging random Fourier features to approximate kernel maps and group-lasso regularisation to enforce sparsity, achieving fixed per-iteration complexity and sublinear dynamic regret guarantees. More recent work tackles the challenge of hidden nodes by assuming smoothness and stationarity of graph signals over partially observed networks; this method jointly estimates successive graph shift operators with column-sparsity constraints to capture edge patterns, demonstrating robustness in synthetic and real-world scenarios where only a subset of nodes is visible.

Graph Learning and Topology Inference Techniques publication trend

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

Technical terms

Graph signal: A set of numerical values associated with the nodes of a graph, representing data inputs or observations.

Graph shift operator: A matrix representation of a graph’s connectivity used to model signal diffusion and neighbourhood relationships.

Sparsity: A property of matrices or models in which only a small number of elements or edges are non-zero, promoting simpler structures.

Stationarity: The condition under which a graph signal’s statistical properties remain invariant under a graph shift operation.

Smoothness: A characteristic of graph signals whereby neighbouring nodes exhibit similar signal values, reflecting gradual variation.

Graph Laplacian: A matrix capturing the difference between a node’s value and the average of its neighbours, encoding connectivity and diffusion dynamics.

References

1. Identifying the Topology of Undirected Networks From Diffused Non-Stationary Graph Signals. IEEE Open Journal of Signal Processing (2021).
2. Learning Time-Varying Graphs From Online Data. IEEE Open Journal of Signal Processing (2022).
3. Sparse Online Learning With Kernels Using Random Features for Estimating Nonlinear Dynamic Graphs. IEEE Transactions on Signal Processing (2023).
4. Time-varying graph learning from smooth and stationary graph signals with hidden nodes. EURASIP Journal on Advances in Signal Processing (2024).