Abstract:Graph signals are functions of the underlying graph. When the edge-weight between a pair of nodes is high, the corresponding signals generally have a higher correlation. As a result, the signals can be represented in terms of a graph-based generative model. The question then arises whether measurements can be obtained on a few nodes and whether the correlation structure between the signals can be used to reconstruct the graph signal on the remaining nodes. We show that node subsampling is always possible for graph signals obtained through a generative model. Further, a method to determine the number of nodes to select is proposed based on the tolerable error. A correlation-based fast greedy algorithm is developed for selecting the nodes. Finally, we verify the proposed method on different deterministic and random graphs, and show that near-perfect reconstruction is possible with node subsampling.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is **how to reconstruct graph signals through node subsampling**, and utilize the correlations between graph signals in this process. Specifically, the author has studied how to select a small number of nodes for measurement given a generative model, and use the signals of these nodes to reconstruct the signals on the entire graph. ### Core of the problem 1. **Correlation of graph signals**: When the edge weight between two nodes in the graph is high, the corresponding graph signals usually have a high correlation. Therefore, graph signals can be represented as a generative model based on the graph structure. 2. **Subsampling possibility**: The question is whether it is possible to obtain measurement values only on a few nodes and use the correlation between signals to reconstruct the graph signals on the remaining nodes. 3. **Error tolerance**: It further explores how many nodes can be ignored when a certain error is allowed. ### Solutions 1. **Generative model**: The author proposes a generative model in which the graph signal \(y(t)\) is a function of graph parameters (especially edge weights). This model is similar to random signal generation, in which white - noise signals generate random signals with specific spectral characteristics through a filter. In this model, the graph plays the role of a filter. 2. **Low - rank approximation**: Due to the low - rank property of the generation matrix \(B\), low - dimensional approximation can be used to reduce the number of nodes that need to be measured. For example, if the weight between two nodes is high, their signals are highly correlated, so the signals of these two nodes can be reconstructed through the signals of other nodes. 3. **Greedy algorithm**: To select the nodes to be measured, the author has developed a correlation - based greedy algorithm. This algorithm selects the least correlated nodes according to the correlation between nodes to ensure that the reconstruction error is minimized. 4. **Reconstruction algorithm**: A reconstruction algorithm is proposed to reconstruct the signals on the entire graph from the measurement values of the selected nodes. ### Experimental verification The author has verified the proposed method on different deterministic and random graphs, and the results show that near - perfect reconstruction can be achieved through subsampling when a certain error is allowed. ### Summary The main contributions of this paper are: - **Theoretical proof**: It is proved that for graph signals generated by the generative model, at least one node can be ignored and the graph signal can still be perfectly reconstructed. - **Practical application**: A correlation - based greedy algorithm is proposed to select the nodes to be measured, and a reconstruction algorithm is designed to recover the signals on the entire graph. - **Experimental verification**: The effectiveness of the method is verified through various types of graphs, showing the node selection and reconstruction performance under different error tolerances. ### Formula summary 1. Generative model of graph signals: \[ y(t)=Bc(t) \] where \(B\) is the generation matrix and \(c(t)\) is the sparse coefficient vector. 2. Measurement model after subsampling: \[ \tilde{y}=\tilde{B}c \] where \(\tilde{B}\) is the low - rank approximation of \(B\). 3. Reconstruction formula: \[ \hat{\tilde{y}}=T(A^T S T)^{-1}y_S \] where \(A_S\) is the subsampling matrix and \(y_S\) is the subsampling - after signal. 4. Reconstruction error: \[ \|y - \hat{\tilde{y}}\|_2=\|(B - \tilde{B})c+T(A^T S T)^{-1}(\tilde{y}_S - y_S)\|_2 \] Through these methods, the author has successfully solved the problem of how to use correlation for subsampling in graph signals and provided theoretical and experimental support.

Subsampling of Correlated Graph Signals

Sensor selection on graphs via data-driven node sub-sampling in network time series

Signal Recovery on Graphs: Fundamental Limits of Sampling Strategies

Generalized Graph Signal Sampling And Reconstruction

Estimating the Number of Connected Components in a Graph via Subgraph Sampling

Sampling and Recovery of Graph Signals based on Graph Neural Networks

Compressive Recovery of Signals Defined on Perturbed Graphs

Sampling and Reconstruction of Signals on Product Graphs

Low-complexity Graph Sampling with Noise and Signal Reconstruction via Neumann Series

Reconstructing Graph Signals from Noisy Dynamical Samples

Learning Graphs from Linear Measurements: Fundamental Trade-offs and Applications

Iterative Reconstruction of Graph Signal in Low-Frequency Subspace.

Signal Recovery on Graphs: Variation Minimization

Kernel-Based Reconstruction of Graph Signals

Graph sub-sampling for divide-and-conquer algorithms in large networks

Time-Varying Graph Signal Reconstruction

Ising on the Graph: Task-specific Graph Subsampling via the Ising Model

Graph Signal Reconstruction from Low-Resolution Multi-Bit Observations

Recovery of Graph Signals from Sign Measurements

Two subspace methods for frequency sparse graph signals

On Critical Sampling of Time-Vertex Graph Signals