Abstract
Identifying stable communities in dynamic networks—such as evolving social circles or fluctuating economic transactions—is a significant challenge due to temporal "jitter," where minor edge changes can lead to drastic shifts in detected community structures. This paper presents MFC-TopoReg, a deep learning framework that integrates Matrix Factorization Clustering (MFC) with Topological Regularization (TopoReg). By utilizing Persistent Homology to analyze the "shape" of inter-community relationships, the framework ensures temporal consistency and filters out noise. Experimental results on datasets like Enron, DBLP, and Highschool demonstrate that MFC-TopoReg significantly outperforms state-of-the-art methods in both accuracy and stability.
Introduction
Community detection aims to group nodes in a network that are more densely connected to each other than to the rest of the graph. In dynamic networks, these connections evolve over time. A major hurdle in this field is temporal inconsistency: most algorithms optimize for the best clustering at individual snapshots, leading to "flickering" results where communities appear to merge or split due to transient noise. For instance, a few temporary interactions between two distinct social groups might cause a standard algorithm to hallucinate a merger. To address this, researchers from Tsinghua University (presented at AAAI-24) propose a solution rooted in Topological Data Analysis (TDA), focusing on the persistent "shape" of the network rather than just local connectivity.
Methods: The MFC-TopoReg Framework
The proposed framework consists of two primary components designed to balance local clustering accuracy with global temporal stability.
Matrix Factorization Clustering (MFC)
Deep graph clustering often suffers from "representation collapse," where neural networks force all nodes in a cluster into a single point, destroying the underlying geometry. MFC avoids this by treating clustering as a dimension-reduction task with relaxed sparsity constraints. This approach preserves the structure of the embedding space while identifying clear group boundaries, preventing the loss of valuable structural information.
Topological Regularization (TopoReg)
The core innovation is TopoReg, which ensures temporal consistency by analyzing the community-level network (a "meta-graph" where nodes represent communities). The framework employs Persistence Homology to generate a "Persistence Barcode"—a topological signature describing the connected components and "holes" in the community structure across different scales.
By calculating the Wasserstein distance between barcodes of consecutive time steps, the model penalizes abrupt changes in the fundamental shape of the communities. Because TDA is scale-independent and robust to small perturbations, it allows the model to ignore temporal noise while remaining sensitive to genuine structural shifts. Crucially, the authors integrated this TDA-based loss into an end-to-end differentiable deep learning pipeline.
Results and Discussion
The model was evaluated on several real-world datasets (Enron, Highschool, DBLP, Cora) and custom synthetic data designed to simulate ephemeral collaborations.
- Accuracy and Stability: On the Highschool dataset, TopoReg improved clustering accuracy from 81% to 98% compared to baselines. Across all datasets, the framework consistently achieved top-tier scores in Accuracy (ACC), Normalized Mutual Information (NMI), and Adjusted Rand Index (ARI).
- Noise Resistance: Synthetic tests confirmed the model's ability to ignore "fake" merging events caused by temporary perturbations that typically trip up baseline algorithms.
- General Improvement: The authors observed an average performance boost of 5.9% to 11.5% across various backbone models when the TopoReg module was added.
Conclusion
MFC-TopoReg represents a significant advancement in dynamic network analysis by bringing the rigors of algebraic topology to deep graph clustering. By focusing on the stability of the inter-community meta-graph, the framework provides a robust "stabilizer" against temporal noise. Future work may explore scaling these TDA methods to massive networks with millions of nodes by utilizing more efficient graph representations.
Summary updated