MEMM: Multilayer Edge Mixture Model

• MEMM is a hyper-model for community detection in multilayer networks that unifies modularity-based and statistical inference methods.
• It decomposes community quality into eight edge and coupling patterns, enabling clear interpretation and flexible parameter tuning.
• The model is optimized with heuristics like the Louvain method and validated through benchmarks showing robust performance and cross-layer consistency.

The Multilayer Edge Mixture Model (MEMM) is a general-purpose hyper-model for community detection in multilayer networks. By representing community quality as a linear combination over eight distinct types of (intra- and interlayer) edge and coupling patterns, MEMM accommodates and generalizes both modularity-based and statistical inference frameworks, providing a robust and interpretable structure for the analysis and evaluation of multilayer community structure (Zhang et al., 2016).

1. Formal Definition and Mathematical Structure

Consider a multilayer network with global node set $V=\{1,\ldots,N\}$ and layers $S=\{1,\ldots,L\}$. Intralayer connectivity is encoded by adjacency matrices $A_{ijs}$, and interlayer couplings by binary variables $C_{isr}$. Community structure can be encoded by hard labels $\upsilon_{is}$ or by soft membership probabilities.

The MEMM evaluator is of the form: $$\begin{aligned} \mathcal M(\upsilon)= &\sum_{s=1}^L\sum_{i\neq j}\left[ \lambda(a)a_{ijs}A_{ijs}P(\upsilon_{is},\upsilon_{js}) + \lambda(b)b_{ijs}(1-A_{ijs})P(\upsilon_{is},\upsilon_{js}) \right.\ &\left. +~ \lambda(c)c_{ijs}A_{ijs}(1-P(\upsilon_{is},\upsilon_{js})) + \lambda(d)d_{ijs}(1-A_{ijs})(1-P(\upsilon_{is},\upsilon_{js})) \right] \ &+\sum_{s\neq r}\sum_{i=1}^N \left[ \lambda(e)e_{isr}C_{isr}P(\upsilon_{is},\upsilon_{ir}) + \lambda(f)f_{isr}(1-C_{isr})P(\upsilon_{is},\upsilon_{ir}) \right.\ &\left. +~ \lambda(g)g_{isr}C_{isr}(1-P(\upsilon_{is},\upsilon_{ir})) + \lambda(h)h_{isr}(1-C_{isr})(1-P(\upsilon_{is},\upsilon_{ir})) \right], \end{aligned}$$ where each hyper-parameter function $\omega \in \{a,b,\ldots,h\}$ is nonnegative, and sign-indicators $\lambda(\omega) \in \{\pm1\}$ determine whether the corresponding pattern is rewarded or penalized.

The probability function $P(\upsilon_{is},\upsilon_{jr})$ encodes either hard cluster membership (as a Kronecker delta) or soft co-membership probabilities.

2. Multilayer Network Modeling and Community Encoding

MEMM assumes a shared node set across layers, with intralayer topologies defined by $A_{ijs}$ and interlayer relationships by $S=\{1,\ldots,L\}$0. Interlayer couplings typically connect node-copies across layers, restricting $S=\{1,\ldots,L\}$1 to identical node indices.

Community structure is modeled in two equivalent forms:

• Hard assignment: $S=\{1,\ldots,L\}$2.
• Soft assignment: $S=\{1,\ldots,L\}$3, the probability of mutual module membership.

The eight hyper-parameter functions distinguish between (internal/external) × (existence/non-existence) for both intralayer edges and interlayer couplings. Adjustment of $S=\{1,\ldots,L\}$4 values controls whether each pattern is encouraged or discouraged by the evaluator.

3. MEMM as a Unification of Modularity and SBMs

By suitably parameterizing $S=\{1,\ldots,L\}$5 through $S=\{1,\ldots,L\}$6 and their signs, MEMM recovers classical multilayer community detection methods as specific cases.

Multilayer Modularity:

With

$S=\{1,\ldots,L\}$7

and parameter choices

$S=\{1,\ldots,L\}$8

where $S=\{1,\ldots,L\}$9 is a null-model probability and $A_{ijs}$0 a resolution parameter, MEMM yields the generalized multilayer modularity of Mucha et al.

Stochastic Blockmodel (SBM) Likelihood:

By setting

$A_{ijs}$1

and similar for coupling terms, MEMM reduces to the multilayer SBM log-likelihood up to a constant, with all $A_{ijs}$2. This demonstrates that MEMM subsumes both modularity-based and likelihood-based multilayer methods (Zhang et al., 2016).

4. Hyper-Parameter Interpretation and Tuning

The eight hyper-parameters control the balance between different edge and coupling types:

Symbol(s)	Edge/Coupling Type	Structural Role
$A_{ijs}$3	Intralayer (internal/external, present/absent)	Governs within-layer structure and resolution $A_{ijs}$4
$A_{ijs}$5	Interlayer (internal/external, present/absent)	Governs interlayer consistency via coupling strength $A_{ijs}$6

• Choice of $A_{ijs}$7 specifies whether a pattern is promoted or suppressed.
• Discriminative weighting: Functions $A_{ijs}$8 chosen to ensure maximal contributions across all types are balanced, enforcing that no single term dominates.
• Resolution parameters $A_{ijs}$9: Small values favor coarse partitions; large values split communities more finely.
• Coupling strength $C_{isr}$0: Controls enforcement of consistent communities across layers. High values force nearly identical assignments.

Parameter selection is often performed by grid search, predictive log-likelihood cross-validation, or by stability analysis under stochastic perturbations.

5. Optimization and Computational Aspects

The maximization of $C_{isr}$1 is NP-hard, paralleling single-layer modularity. Practical optimization uses a multilayer Louvain-style greedy heuristic. Each iteration consists of:

• Local moves: Relocating individual node-copies to maximize local $C_{isr}$2 gain.
• Aggregation: Reducing communities to super-nodes and repeating.

Computational complexity per iteration is $C_{isr}$3 with $C_{isr}$4 as the total edge and coupling count. Empirically, a small number of passes suffice. For SBM-based instantiations, an expectation–maximization scheme on $C_{isr}$5 is employed.

6. Empirical Performance and Benchmark Analysis

Experiments used synthetic multilayer benchmarks:

• Normal four-layer LFR graphs ($C_{isr}$6 nodes/layer, degree $C_{isr}$7, four equal modules).
• Four-layer bipartite planted-partition models with $C_{isr}$8.

Key findings:

• Rewarding scheme: Standard modularity assignment yields high normalized mutual information (NMI) on conventional layers, fails on bipartite; sign-swapped (“bipartite modularity”) inverts this pattern. MEMM's explicit edge-type decomposition clarifies these tendencies.
• Coupling weighting: Counting both existing and non-existing interlayer couplings produces more robust NMI, except a notable instability near coupling density $C_{isr}$9, explained by net coupling contribution vanishing.
• Parameter sweeps: On multilayer Zachary karate club replications ($\upsilon_{is}$0 layers, varying $\upsilon_{is}$1), stronger coupling ($\upsilon_{is}$2) induces cross-layer assignment alignment but can override resolution distinctions; adjustment of $\upsilon_{is}$3 tunes module granularity as predicted.

7. Flexibility, Theoretical Scope, and Limitations

MEMM unifies a broad range of community detection strategies in multilayer settings:

• Supports modularity-like and likelihood-based evaluators via parametric specialization.
• Decomposes any multilayer quality function into interpretable contributions from eight structural patterns.
• Permits the creation of new measures by adjusting or designing hyper-parameter functions ($\upsilon_{is}$4).

However, the model’s expressive power is offset by practical challenges:

• The space of eight hyper-parameters plus sign settings is high-dimensional; improper scaling yields degenerate or trivial solutions.
• Like modularity, MEMM optimization is non-convex and subject to local optima and “resolution limit” effects.
• Special care is required in coupling-dense regimes; when both existing and non-existing couplings are equally weighted, the evaluator can lose discriminatory power at $\upsilon_{is}$5.

Overall, MEMM provides a principled and extensible foundation for multilayer community detection, contingent on informed parameterization and robust optimization strategies (Zhang et al., 2016).