TL;DR
This paper introduces the concept of "local stationarity" for graph signals, extending the existing notion of global stationarity. By defining a "local power spectrum," the authors provide a new way to pinpoint sources of non-stationarity within a graph signal and develop a more robust power spectrum estimator.
What Problem Does This Paper Address?
Graph signal processing (GSP) aims to analyze data structured as graphs. A key concept in traditional signal processing is stationarity, meaning a signal's statistical properties don't change over time. Extending this to graphs, "graph stationarity" implies statistical invariance under a "graph translation" operator. While this global definition is useful, it offers limited insight into where within the graph a signal might be non-stationary. The authors aim to bridge this gap by developing a concept of local stationarity, allowing for a more granular understanding of signal behavior across different parts of the graph. This is crucial for applications where localized anomalies or variations are of interest.
Methods
The core of the paper is the introduction of the local power spectrum. This is achieved by:
- Localized Atoms: The authors define a set of "localized atoms" (functions that are concentrated around specific vertices and have varying "spans" or scales). These atoms are constructed using a novel "Expo" basis, designed to avoid biases in the frequency domain, and are compared to atoms from the Spectral Graph Wavelet Transform (SGWT).
- Local Power Spectrum Calculation: For a given vertex i and scale m, the local power spectrum $S_x(i, m)$ is defined as the expected squared magnitude of the signal's projection onto the localized atom centered at i with span m.
- Non-stationarity Detection: By examining how the local power spectrum varies across different vertices, the authors can identify regions of the graph where the signal's statistical properties differ significantly.
- Power Spectrum Estimation: For stationary graph signals, the sum of local power spectra across all vertices can be used to estimate the global power spectrum. This approach allows for estimation even when using only a subset of the localized atoms (i.e., truncated decomposition).
Datasets
- Molène Graph Dataset: This dataset consists of hourly weather readings from 28 weather stations in the Brittany region of France, collected over January 2014. The graph is constructed using a Gaussian kernel based on geographical distances between stations.
- Synthetic Data (GMRF Model): The authors generate 744 realizations of a Gaussian Markov Random Field (GMRF) model with a smooth power spectrum. This synthetic data is used to validate the proposed methods when the ground truth is known.
Key Insights and Ideas
- Vertex-Domain Stationarity: The concept of local stationarity provides a much-needed interpretation of stationarity directly in the vertex domain, making it more intuitive than the global definition.
- Pinpointing Non-Stationarity: The local power spectrum acts like a "heat map" for signal anomalies, clearly highlighting specific vertices or regions that deviate from stationary behavior. This is a significant improvement over global methods that can obscure localized issues.
- Atom Choice Matters: The choice of localized atoms (e.g., SGWT vs. Expo) impacts the stability and accuracy of both non-stationarity detection and power spectrum estimation, particularly when dealing with truncated decompositions.
- Global Stationarity Doesn't Imply Local Stationarity: The paper demonstrates that a globally stationary signal doesn't necessarily exhibit uniform local properties, underscoring the value of the local approach.
Key Results
- Non-stationarity Identification: On the Molène weather data, the local power spectrum successfully identified a specific weather station (vertex #19) exhibiting significantly different spectral behavior compared to others. This deviation was further analyzed and attributed to rapid temperature fluctuations at that location.
- Power Spectrum Estimation: The proposed local power spectrum-based estimator, when applied to synthetic GMRF data, performed comparably to existing methods, with the SGWT-based approach showing more stable results when using truncated decompositions.
- Comparison of Estimators: The paper compares its estimator to a "Simple" estimator (direct Fourier transform and averaging) and the Perraudin et al. method, showing competitive performance, particularly in capturing spectral richness at lower frequencies.
Limitations and Future Work
- Atom Design: The authors acknowledge that the formal definition and optimal selection of the set of localized atoms $\{g_{i,m}\}$ is an ongoing area of research. Future work will focus on developing better atoms for improved localization and power spectrum estimation.
- Complex Eigenvalue Cases: The current work primarily addresses graphs with distinct eigenvalues. Extending the framework to handle graphs with multiple eigenvalues is a planned future direction.
- Local Stationarity Test: The authors are looking into developing a formal statistical test for local stationarity, building upon existing work.
- Broader Applicability: Investigating how this framework performs for different classes of stochastic graph signals is also a future research avenue.
Summary updated