Degree-Corrected Joint Matrix Factorization for Multilayer Community Detection
This paper tackles community detection in multilayer networks where connectivity patterns and node degrees vary across layers. The proposed method, mFROST, jointly factorizes each layer via orthogonal symmetric nonnegative matrix trifactorization $A_l \approx D_l V \theta_l (D_l V)^\top$, enforcing shared community assignments $V$ while allowing layer-specific degree corrections $D_l$ and community interactions $\theta_l$. This flexibility matters because real-world multiplex networks often exhibit heterogeneous node sociability and diverse block structures that simpler shared-factor models miss.
mFROST introduces a well-motivated and algorithmically efficient extension of orthogonal symmetric nonnegative matrix trifactorization to the multilayer setting. Its close correspondence to the multilayer degree-corrected stochastic block model yields interpretable parameters, and the block coordinate descent scheme scales linearly with the total number of edges. However, the paper does not establish statistical consistency or recovery theory under that generative model, and the squared Frobenius loss ignores the Bernoulli nature of adjacency entries, which can lead to heteroskedasticity issues. The method is also limited to disjoint communities and remains outperformed by Bayesian inference on some real-world networks.
The strongest aspect of the paper is its principled modeling choice to separate shared node memberships from layer-specific sociability and community interaction matrices. The ablation study demonstrates that allowing $Z_l$ to vary across layers significantly improves recovery when node degrees differ across layers, without degrading performance when degrees are identical. The optimization procedure is carefully derived, with exact closed-form updates for $\theta_l$ and an efficient polynomial solver for the row-wise updates of $Z_l$, supported by a theoretically grounded initialization strategy from the authors' prior work.
The objective relies on the sum of squared Frobenius reconstruction errors across layers, which treats binary adjacency entries as continuous Gaussian observations and does not account for heteroskedastic variance or sparsity-dependent noise. Despite the close connection to the MDCBM, the paper offers no theoretical guarantees for exact or approximate recovery under that generative model. In synthetic regimes where layers are homogeneous and few in number, the simpler spectral baseline Sum A outperforms mFROST, suggesting that the added parameter flexibility requires sufficient layers to pay off. The restriction to disjoint communities and the reliance on multiple random restarts to handle non-convexity are further practical limitations.
The empirical evaluation is thorough, reproducing the synthetic benchmark from Agterberg et al. and testing on four real-world datasets against spectral, Bayesian, and matrix-factorization competitors. The ablation isolating layer-specific $Z_l$ is particularly convincing. That said, the real-world results are mixed: graph-tool achieves higher NMI on Cora and AUCS, and the advantage of mFROST is most pronounced on synthetic networks with heterogeneous degree or interaction patterns. The authors honestly note that graph-tool is the strongest competitor overall, tempering claims of universal dominance.
Reproducibility is aided by a public GitHub repository and explicit reporting of the stopping tolerance ($\delta=10^{-5}$), maximum iterations, and number of restarts. The initialization strategy is detailed, and Table III documents fast empirical convergence across datasets. However, randomness in the initialization procedure and the lack of reported random seeds mean that exact reproduction of the best selected restart may vary across machines. Furthermore, baseline comparisons depend largely on default hyperparameters imported from the DC-MASE reference implementation, so it is unclear whether competitor methods were tuned to the same extent as mFROST.
Multilayer networks allow the modeling of interactions between the same entities across different contexts, such as temporal observations, varying settings, or interactions of different types. The goal of community detection in multilayer networks is to identify groups of nodes exhibiting similar connectivity patterns, which may vary across layers. We propose a method based on a joint nonnegative symmetric matrix trifactorization for community detection in multilayer networks, where each graph is approximated by a nonnegative symmetric matrix trifactorization. Our approach enforces constraints on the factor matrices so that communities are disjoint and shared across layers, while allowing each layer to have its own connectivity patterns and node degrees. This flexibility enables the model to capture both local and global structural variations across layers. We also develop an algorithm to efficiently solve this problem. We evaluate multilayer community detection methods using the multilayer degree-corrected stochastic block model (MDCBM), a flexible framework for generating realistic multilayer graphs with heterogeneous degrees and varying connectivity patterns. Experiments show that our method reliably detects communities across diverse regimes, whereas existing state-of-the-art approaches are often limited by restrictive structural assumptions.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.