TL;DR
This paper introduces a more principled, parameter-free approach to measuring modularity in multilayer networks. By replacing arbitrary global constants with data-driven factors—specifically a redundancy-based resolution and projection-based inter-layer coupling—the authors provide a more expressive metric that naturally accounts for community-specific structures and temporal layer orderings.
What Problem Does This Paper Address?
In network science, "modularity" is the standard yardstick for measuring the strength of a network’s division into communities. While well-defined for simple graphs, multilayer networks (where nodes are connected via multiple types of relations, like "friendship" on one layer and "co-authorship" on another) present a challenge.
The existing standard, multislice modularity ($Q_{ms}$), relies on two crucial parameters: a resolution factor ($\gamma$) and an inter-layer coupling factor ($\omega$). Traditionally, these are set as arbitrary constants. This is problematic because it ignores the unique topological characteristics of different layers and communities. Furthermore, $Q_{ms}$ doesn't inherently handle "ordered" layers, such as those representing a network evolving over time, where a node's state in one time-step should primarily be coupled with its state in the next.
Methods
The authors propose a new formulation of multilayer modularity ($Q$) that eliminates the need for manual parameter tuning:
- Redundancy-based Resolution ($\gamma$): Instead of a fixed value, $\gamma$ is calculated for each community-layer pair. It uses the concept of "redundancy"—the variety of link types between node pairs. If a community has high information content (many redundant links across layers), the resolution factor is lowered to reflect that layer's high contribution to the community's internal structure.
- Projection-based Inter-Layer Coupling ($IC$): Instead of a constant $\omega$, the coupling strength is determined by how much a community’s "projection" (the set of nodes it contains) overlaps between two layers.
- Asymmetry and Ordering: The model introduces asymmetric coupling, recognizing that a community's relevance might differ when projecting from a "source" layer to a "destination" layer. It also supports partial order relations, allowing researchers to specify if layers are adjacent (only coupling $L_i$ to $L_{i+1}$) or succeeding (coupling $L_i$ to all $L_j$ where $j > i$).
- Time-Awareness: For temporal networks, they include a smoothing factor that penalizes the coupling strength as the distance between layers (time-steps) increases.
Datasets
The authors validated their approach using a mix of synthetic and real-world data:
- Synthetic: They generated networks using Erdős-Rényi (ER) and Lancichinetti-Fortunato-Radicchi (LFR) benchmarks, creating various multilayer configurations (e.g., ER-ER, LFR-ER) to test how the metric responds to known community structures.
- Real-World: 10 diverse datasets were used, including:
- AUCS: University employee relationships (work, lunch, Facebook, etc.).
- EU-Air: European airport connections across different airlines.
- Social/Digital Networks: FriendFeed-Twitter-YouTube (FF-TW-YT), Flickr, GitHub-StackOverflow-Twitter (GH-SO-TW), and Higgs-Twitter.
- Transport/Other: London transport network, Obama’s 2013 Israel visit (Twitter interactions), and a classroom social network (7thGraders).
Key Insights and Ideas
- Context Matters: Modularity shouldn't be a "one-size-fits-all" global constant. By making $\gamma$ and $IC$ community-specific, the metric becomes much more sensitive to the actual "meaningfulness" of a group of nodes.
- The "Surprise" of Projections: The asymmetric coupling idea is particularly clever. It captures the intuition that a small, tightly-knit community in one layer projecting perfectly onto a larger layer is more "surprising" (and thus indicates stronger coupling) than a giant community projecting onto a subset of itself.
- Temporal Logic: By integrating layer ordering directly into the modularity equation, the authors move away from the "all-to-all" coupling of previous models, which often led to overestimating community stability in time-evolving networks.
Key Results
- Higher Expressiveness: The proposed $Q$ consistently showed higher values and better differentiation than the standard $Q_{ms}$ across synthetic benchmarks.
- Redundancy Gains: Using the redundancy-based $\gamma$ led to modularity gains of up to 0.12 compared to fixing $\gamma=1$, proving that data-driven resolution captures more structural information.
- Asymmetry Wins: Asymmetric coupling ($IC_a$) generally resulted in higher modularity scores than symmetric versions, supporting the idea that directional relevance is a key feature of multilayer systems.
- Structural Correlation: The authors found a strong positive correlation between their modularity measure and key topological features like edge-set coverage, suggesting the metric is a reliable proxy for "true" community quality.
Limitations and Future Work
- Computational Overhead: The authors note that calculating variable $\gamma$ and $IC$ factors increases computation time (roughly a 50% increase in their LFR tests). While still efficient for medium networks, very large-scale multiplex graphs might require optimization.
- Crisp vs. Overlapping: The current model focuses on "crisp" community memberships (a node is either in or out). The authors suggest that extending this to overlapping communities (soft memberships) is a vital next step.
- Optimization Algorithms: This paper defines the measure of modularity. The next logical phase is developing new community detection algorithms specifically designed to maximize this new version of $Q$.
Reference: Amelio, A., Mangioni, G., & Tagarelli, A. (2019). Modularity in Multilayer Networks using Redundancy-based Resolution and Projection-based Inter-Layer Coupling. IEEE Transactions on Network Science and Engineering. DOI: 10.1109/TNSE.2019.2913325.
Summary updated