CLC-NMF: Cross-Layer Community Geometry Transfer for Link Prediction in Multilayer Networks

A unified research framework for reproducible experiments on multilayer network link prediction using Community Geometry Transfer.

---

Overview

CLC-NMF is a lightweight, training-free framework for improving intra-layer link prediction in heterogeneous multilayer networks.

Instead of transferring node embeddings or requiring node correspondence across layers, CLC-NMF transfers community geometry, represented by the Gram Matrix of Non-negative Matrix Factorization (NMF) latent factors.

The framework introduces:

• Reliability-weighted knowledge transfer
• Community geometry alignment
• Validation-gated transfer to minimize negative transfer
• A reproducible experimental pipeline for benchmarking multilayer link prediction methods

---

Research Objectives

This project aims to answer the following research questions:

• Can cross-layer community geometry improve link prediction?
• Which network layers benefit most from knowledge transfer?
• Can reconstruction reliability identify trustworthy source layers?
• How can negative transfer be reduced?
• How robust is community geometry transfer across different datasets?

---

Repository Structure

CLC-NMF/
│
├── datasets/              # Benchmark multilayer datasets
│
├── methods/               # Link prediction algorithms
│   ├── standard_nmf.py
│   ├── clc_nmf.py
│   ├── common_neighbors.py
│   ├── adamic_adar.py
│   ├── jaccard.py
│   ├── node2vec.py
│   └── deepwalk.py
│
├── experiments/
│   ├── configs/
│   ├── run_experiment.py
│   ├── evaluator.py
│   ├── statistics.py
│   ├── logger.py
│   └── plots.py
│
├── outputs/
│   ├── raw/
│   ├── processed/
│   ├── figures/
│   └── tables/
│
├── paper/
│
└── README.md

---

Experimental Pipeline

The entire project is built around a single configurable experiment pipeline.

Every experiment is defined by:

• Dataset
• Method
• Hyperparameters
• Random Seed
• Evaluation Metrics

The pipeline automatically:

1. Loads datasets
2. Performs train/validation/test splitting
3. Trains the selected method
4. Generates link prediction scores
5. Computes evaluation metrics
6. Performs statistical analysis
7. Logs results
8. Generates publication-ready figures and tables

---

Supported Methods

Current

• Standard NMF
• CLC-NMF
• Common Neighbours
• Adamic-Adar
• Jaccard Coefficient

Planned

• Node2Vec
• DeepWalk
• Variational Graph Autoencoder (VGAE)
• Graph Autoencoder (GAE)
• Additional graph embedding baselines

---

Supported Datasets

Current benchmark datasets include:

• London Transport
• Drosophila
• C. elegans
• CKM Physicians
• Kapferer Tailor Shop
• Vickers Classroom Network

The framework is designed to allow easy integration of additional multilayer datasets.

---

Evaluation Metrics

Current

• ROC-AUC

Planned

• Average Precision (AP)
• Precision@K
• Recall@K
• F1 Score
• PR-AUC
• Runtime
• Memory Usage

---

Statistical Analysis

The framework supports reproducible statistical evaluation.

Current analyses:

• Mean
• Standard Deviation
• 95% Confidence Interval
• Wilcoxon Signed-Rank Test

Planned analyses:

• Cohen's d
• Cliff's Delta
• Bootstrap Confidence Intervals

---

Experimental Roadmap

Phase 1 — Experimental Validation

• Implement stronger baselines
• Add additional evaluation metrics
• Statistical significance testing
• Benchmark comparison

Phase 2 — Robustness Analysis

• Hyperparameter sensitivity
• Runtime scaling
• Memory profiling
• Reliability analysis
• Correlation studies

Phase 3 — Ablation Studies

• Remove reliability weighting
• Remove validation gate
• Remove Gram normalization
• Remove community alignment
• Random transfer
• Density-based weighting
• Equal weighting

Phase 4 — Method Improvements

• Adaptive coupling coefficient
• Similarity-aware transfer
• Improved reliability estimation
• Structural compatibility modeling

---

Experiment Registry

Every experiment is tracked using a unique identifier.

ID	Research Question	Status
E1	Baseline comparison	Planned
E2	Alpha sensitivity	Planned
E3	Latent dimension sensitivity	Planned
E4	Runtime analysis	Planned
E5	Reliability correlation	Planned
E6	Ablation study	Planned
E7	Modern baseline comparison	Planned

---

Reproducibility

Every experiment automatically records:

• Dataset
• Method
• Hyperparameters
• Random Seed
• Metrics
• Runtime
• Memory Usage

This enables complete regeneration of every figure, table, and statistical result presented in the accompanying paper.

---

Current Development Focus

The project is currently focused on:

• Building a unified experimental pipeline
• Strengthening empirical evaluation
• Adding modern baseline methods
• Improving statistical validation
• Extending ablation studies
• Enhancing reproducibility

---

Contributing

Contributions are welcome.

If you would like to improve the framework, implement additional methods, or add new datasets, feel free to open an issue or submit a pull request.

---

License

This project is intended for academic and research purposes.

A formal open-source license will be added upon public release.

---

Citation

If you use this repository in your research, please cite the accompanying publication (citation will be added after publication).