# Deep-spectrum community features

- ID: 2121
- Canonical URL: https://synthcore.org/idea/2121/deep-spectrum-community-features
- API JSON: https://synthcore.org/api/idea/2121.json
- API Markdown: https://synthcore.org/api/idea/2121.md
- Verification status: unverified
- Source: [arXiv:2608.11321](https://arxiv.org/abs/2608.11321)
- Category: architecture
- Solves: accuracy, stability, sample-efficiency
- ML areas: graph-nn, embedding, attention
- Math tags: spectral-theory, graph-theory, probability, statistics
- Ratings: usefulness 6/10; difficulty 5/10; novelty 6/10

## Idea description

Replace the usual top-eigenvector positional encoding in a graph neural network with a density-selected spectral subspace. The selector explicitly searches below the leading eigenvectors, where community information may survive after latent geometric modes have consumed the largest eigenvalues. The selected coordinates can be concatenated to node features or used as a bias in graph attention.

## Mathematical statement

The model defines a symmetric label-interaction matrix $\mathbf P=\begin{pmatrix}a&b\\b&a\end{pmatrix}$, where $a$ is the same-cluster connection coefficient and $b$ is the different-cluster coefficient. For latent positions $X_i,X_j$ and cluster labels $\sigma_i,\sigma_j\in\{1,2\}$, the conditional edge probability is $\mathbb P(A_{ij}=1\mid X,\sigma)=\rho_n K(X_i,X_j)\mathbf P(\sigma_i,\sigma_j)$, where $K$ is a symmetric measurable latent-geometry kernel and $\rho_n$ controls sparsity. The matrix decomposes as $\mathbf P=\frac{a+b}{2}\mathbf 1\mathbf 1^\top+\frac{a-b}{2}\begin{pmatrix}1&-1\\-1&1\end{pmatrix}$, with $\mathbf 1=(1,1)^\top$. The second term is the community contrast, while the kernel $K$ can create larger eigenmodes associated with latent geometry. Consequently, the community eigenvector may occur deeper in the spectrum rather than among the leading eigenvectors. The adaptation computes eigenpairs $(\lambda_r,u_r)$ of a regularized normalized adjacency matrix, forms candidate spectral windows that omit the first few eigenvectors, and selects the window with the strongest density-based cluster structure instead of assuming that the largest eigenvalues are informative.

## Key formulas

- $$\mathbf P=\begin{pmatrix}a&b\\b&a\end{pmatrix},\qquad \mathbb P(A_{ij}=1\mid X,\sigma)=\rho_n K(X_i,X_j)\mathbf P(\sigma_i,\sigma_j).$$
- $$\mathbf P=\frac{a+b}{2}\mathbf 1\mathbf 1^\top+\frac{a-b}{2}\begin{pmatrix}1&-1\\-1&1\end{pmatrix},\qquad \mathbf 1=(1,1)^\top.$$
- $$\widetilde A=D_\tau^{-1/2}(A+\tau I)D_\tau^{-1/2},\qquad D_\tau=\operatorname{diag}\bigl((A+\tau I)\mathbf 1_n\bigr).$$
- $$U_{s,w}=[u_s,u_{s+1},\ldots,u_{s+w-1}],\qquad (s^*,w^*)=\arg\max_{s,w}\operatorname{Silhouette}\!\left(\operatorname{DBSCAN}(U_{s,w})\right).$$

## Implementation notes

Integrate this at the input of a GNN or in the key/query computation of a graph-attention layer. Given an undirected adjacency matrix $A\in\{0,1\}^{n\times n}$, node features $H\in\mathbb R^{n\times d}$, regularization $\tau$, and maximum spectral rank $R$, form $A_\tau=A+\tau I$ and $D_\tau=\operatorname{diag}(A_\tau\mathbf 1_n)$. Start with $\tau$ equal to the mean degree and compute the first $R$ eigenpairs of $\widetilde A=D_\tau^{-1/2}A_\tau D_\tau^{-1/2}$ using Lanczos. Sort eigenpairs by decreasing eigenvalue. For each discard count $q\in\{1,2,4\}$ and each contiguous window width $w\in\{1,2,4,8\}$, construct $U_{s,w}=[u_s,\ldots,u_{s+w-1}]$ for $s>q$, row-normalize its rows, and run DBSCAN. Set the DBSCAN radius from the 10th, 20th, and 30th percentiles of each node's eighth-nearest-neighbor distance, and use minimum sample count 10. Score every candidate by silhouette coefficient, ignoring noise points and penalizing excessive cluster counts when the expected count is known. Select the best $Z=U_{s^*,w^*}$ and concatenate it with $H$, or add an attention-logit bias $B_{ij}=\alpha z_i^\top z_j$ before softmax. The paper-derived quantities are the normalized adjacency, deeper spectral windows, and density-based selection; $q$, DBSCAN parameters, and $\alpha$ are estimated empirically. First test on synthetic two-community graphs with latent points on a circle or in a two-dimensional Gaussian geometry and edge probabilities matching $\rho_nK(X_i,X_j)\mathbf P(\sigma_i,\sigma_j)$. Compare a GCN or graph transformer using standard top-$k$ Laplacian positional encodings against this selector at equal parameter count. Success means higher node-classification accuracy when geometric modes dominate, reduced variance across random graph realizations, and a selected window that moves deeper as the latent-geometry amplitude increases.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.