Abstract:Complex time-varying networks are prominent models for a wide variety of spatiotemporal phenomena. The functioning of networks depends crucially on their connectivity, yet reliable techniques for determining communities in spacetime networks remain elusive. We adapt successful spectral techniques from continuous-time dynamics on manifolds to the graph setting to fill this gap. We formulate an inflated dynamic Laplacian for graphs and develop a spectral theory to underpin the corresponding algorithmic realisations. We develop spectral clustering approaches for both multiplex and non-multiplex networks, based on the eigenvectors of the inflated dynamic Laplacian and specialised Sparse EigenBasis Approximation (SEBA) post-processing of these eigenvectors. We demonstrate that our approach can outperform the Leiden algorithm applied both in spacetime and layer-by-layer, and we analyse voting data from the US senate (where senators come and go as congresses evolve) to quantify increasing polarisation in time.
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is to provide a reliable technique for determining the community structure in spatio - temporal networks, especially those complex networks that evolve over time. Specifically, the authors introduced an inflated dynamic Laplacian and developed the corresponding spectral theory to support algorithm implementation. This method can be used for multi - layer and non - multi - layer networks, perform spectral clustering based on the eigenvectors of the inflated dynamic Laplacian, and post - process these eigenvectors in combination with Sparse EigenBasis Approximation (SEBA).
### Main Problems and Solutions
1. **Problem Description**:
- **Community Detection in Spatio - Temporal Networks**: Traditional community detection methods are not effective in dealing with networks that change over time, especially when the community structure evolves over time.
- **Limitations of Existing Methods**: Existing non - spatio - temporal and spatio - temporal methods have deficiencies in dealing with complex time - evolving networks and cannot effectively capture the dynamic changes of communities.
2. **Proposed Solutions**:
- **Inflated Dynamic Laplacian**: The authors applied the successful spectral techniques in continuous - time dynamical systems to the graph setting and proposed the inflated dynamic Laplacian. This matrix better captures the dynamic characteristics of spatio - temporal networks by duplicating vertices and expanding the graph structure over time.
- **Spectral Theory and Algorithm**: Based on the inflated dynamic Laplacian, the authors developed a new spectral theory and proposed a robust spectral clustering algorithm. This algorithm can detect multiple spatio - temporal communities in multi - layer and non - multi - layer networks and is, in some cases, superior to the existing Leiden algorithm.
- **Application Example**: Through the analysis of the US Senate voting data, it is shown that this method can quantify the change of political polarization over time.
### Formula Representation
- **Inflated Dynamic Laplacian**:
\[
L(a)=\text{blockdiag}(L(1),\ldots,L(5)) + a^{2}(L_{\text{temp}}\otimes I)
\]
where \(L(i)\) is the unnormalized Laplacian matrix of the \(i\)-th time slice, \(L_{\text{temp}}\) is the Laplacian matrix in the time dimension, \(a\) is a scalar parameter used to balance the contributions of the time and space components, \(\otimes\) represents the Kronecker product, and \(I\) is the identity matrix.
- **Spatio - Temporal Cut - Value**:
\[
\sigma(X)=\sum_{(t,x)\in X}\sum_{(s,y)\in X^{c}}W^{(a)}_{(t,x),(s,y)}
\]
where \(X\) is a subset of the vertex set, \(X^{c}\) is its complement, and \(W^{(a)}\) is the inflated dynamic weight matrix.
- **Cheeger Inequality**:
\[
h_{2}\leq\sqrt{2\lambda_{2,a}}
\]
where \(h_{2}\) is the normalized Cheeger constant and \(\lambda_{2,a}\) is the second eigenvalue of the normalized inflated dynamic Laplacian matrix.
Through these methods and formulas, the authors provide an effective framework for detecting and analyzing the community structure in spatio - temporal networks, especially when the community evolves over time.