Spectral Triadic Decompositions of Real-World Networks

TL;DR

This work addresses the challenge of uncovering rich community structure in real-world networks by introducing the spectral triadic decomposition, a framework built around the spectral transitivity $\tau(G)$ and triangle-weighted graph statistics. It proves a central theorem: when $\tau$ is constant, the graph can be decomposed into disjoint, dense, uniformly structured blocks $X_i$ whose internal submatrices $\mathcal{A}|_{X_i}$ are strongly $\mathrm{poly}(\tau)$-uniform and which collectively capture a constant fraction of the graph’s Frobenius norm; the decomposition is computable via a triangle-enumeration-based algorithm. Empirically, the method yields many dense, moderately sized clusters that correlate with semantically meaningful communities in social, coauthorship, and citation networks, often outperforming standard baselines in density and semantic coherence. The results offer a new spectral-graph-theory paradigm for analyzing triangle-rich networks, with robustness to noise and no requirement to pre-specify the number of communities, potentially improving practical community detection in large-scale networks. Overall, the paper provides both a rigorous decomposition theorem and practical algorithmic and empirical evidence that triangle-driven spectral structure underlies real-world community organization.

Abstract

A fundamental problem in mathematics and network analysis is to find conditions under which a graph can be partitioned into smaller pieces. The most important tool for this partitioning is the Fiedler vector or discrete Cheeger inequality. These results relate the graph spectrum (eigenvalues of the normalized adjacency matrix) to the ability to break a graph into two pieces, with few edge deletions. An entire subfield of mathematics, called spectral graph theory, has emerged from these results. Yet these results do not say anything about the rich community structure exhibited by real-world networks, which typically have a significant fraction of edges contained in numerous densely clustered blocks. Inspired by the properties of real-world networks, we discover a new spectral condition that relates eigenvalue powers to a network decomposition into densely clustered blocks. We call this the \emph{spectral triadic decomposition}. Our relationship exactly predicts the existence of community structure, as commonly seen in real networked data. Our proof provides an efficient algorithm to produce the spectral triadic decomposition. We observe on numerous social, coauthorship, and citation network datasets that these decompositions have significant correlation with semantically meaningful communities.

Figures (7)

• Figure 1: On the left, as a small example, we consider a subgraph induced by $155$ vertices and $\tau=0.49$ from a coauthorship network of Condensed Matter Physics researchersNewmanCondMat99, and show a spectral triadic decomposition of the largest connected component, which has $49$ vertices. Each cluster is colored differently. We see how each cluster forms a densely connected component within an otherwise sparse graph. Also note that the clusters vary in size. The gray vertices do not participate in the decomposition, since they do not add significant to the cluster structure. On the right, we look at the adjacency matrices pre and post decomposition. The top figure is a spy plot of the adjacency matrix of 488 connected vertices from a Facebook network (traud2012social,Traud:2011fs) taken from the network repositorynr, a graph with $\tau=0.122$. As a demonstration, we compute the spectral triadic decomposition of this subnetwork. We group the columns/rows by the clusters in the spy plot on the bottom. The latent community structure is immediately visible. Note that there exists many such blocks of varying sizes.
• Figure 2: Mean edge density, triangle density, and fraction of vertices preserved in clusters extracted from caHepTh for $\varepsilon=0.1, 0.2, 0.3, 0.5$. Edge and triangle density rise with increasing $\varepsilon$, but fewer vertices are preserved.
• Figure 3: We give scatter plots for size vs edge density of the spectral triadic clusters and Louvain clusters. In all cases, spectral triadic clusters are dense and of moderate size. While many Louvain cluster are also dense, it also creates a few, extremely large, sparse clusters. Overall, the spectral triadic scatterplot is above the corresponding Louvain plot, though there is significant overlap.
• Figure 4: Uniformity across clusters in the decomposition obtained from various networks as labelled.
• Figure 5: We show two example clusters from a spectral triadic decomposition of coauthorship network of researchers in Condensed Matter Physics NewmanCondMat99, a graph with $\tau=0.25$. The left cluster is a set of $16$ researchers ($58$ edges) working on optics and Bose-Einstein condensates (notably, the cluster has the 2001 Physics Nobel laureate Wolgang Ketterle). The right cluster has $18$ researchers ($55$ edges) working on nanomaterials, including the 1996 Chemistry Nobel laureate Richard Smalley.

Theorems & Definitions (30)

• Definition 1.1
• Definition 1.1
• Theorem 1.2: Spectral Theorem
• Definition 3.1
• Claim 3.2
• proof
• Claim 3.3
• proof
• Claim 3.4
• proof