Research map / Probabilistic and scientific modeling / Molecules and proteins
REM: A Scalable Reinforced Multi-Expert Framework for Multiplex Influence Maximization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes REM, a VAE-based seed-set generator (Seed2Vec) trained as an RL policy with a mixture-of-GNN-experts propagation predictor (PMoE) to maximize influence spread in multiplex networks.
Problem
Multiplex Influence Maximization: selecting a seed set of bounded size in multi-layer social networks to maximize total activated nodes, where prior learning methods struggle with generalization to unknown diffusion patterns, dependence on training data quality, and inaccurate/unscalable spread esti
Approach
Encodes seed sets into a continuous VAE latent space, explores that space with gradient-based RL-style search guided by a mixture-of-GNN-experts spread predictor, and iteratively retrains the VAE on top-ranked generated seed sets combined with the original data.
Key results
- Under IC with 20% budget, REM reaches 1404.14 on Cora-ML vs. 1368.56 for ISF-based baseline sets and achieves best results on NYClimateMarch2014 and ParisAttack2015 among compared methods (e.g., 637621.59 on ParisAttack2015 at 20%).
- Under LT with 20% budget, REM is reported to be about 10% and 15% higher in influence spread than the best competitors on NYClimateMarch2014 and ParisAttack2015.
- Inference runtime at 50,000 nodes is 33.482s for REM vs. 36.437s for MIM-Reasoner (second fastest), though MIM-Reasoner is faster at 10,000 nodes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use REM instead of the strongest baseline ISF | REM · Cora-ML, IC, 20% budget | influence spread: 1368.56 -> 1404.14 | Helped |
| use REM instead of the strongest baseline ISF | REM · Celegans, IC, 20% budget | influence spread: 2819.26 -> 2904.03 | Helped |
| use REM instead of the strongest baseline ISF | REM · Arabidopsis, IC, 20% budget | influence spread: 4694.16 -> 4701.75 | Helped |
| use REM instead of the strongest baseline MIM-Reasoner | REM · NYClimateMarch2014, IC, 20% budget | influence spread: 21062.87 -> 23142.83 | Helped |
| use REM instead of the strongest baseline MIM-Reasoner | REM · ParisAttack2015, IC, 20% budget | influence spread: 608192.57 -> 637621.59 | Helped |
| use REM instead of the strongest baseline ISF | REM · Cora-ML, LT, 20% budget | influence spread: 2145.0 -> 2141.0 | Hurt |
| use REM instead of the strongest baseline ISF | REM · Celegans, LT, 20% budget | influence spread: 3857.0 -> 3877.0 | Helped |
| use REM instead of the strongest baseline ISF | REM · Arabidopsis, LT, 20% budget | influence spread: 6855.0 -> 6890.0 | Helped |
| use REM instead of the strongest baseline MIM-Reasoner | REM · NYClimateMarch2014, LT, 20% budget | influence spread: 96239.0 -> 98976.0 | Helped |
| use REM instead of the strongest baseline MIM-Reasoner | REM · ParisAttack2015, LT, 20% budget | influence spread: 786819.0 -> 804469.0 | Helped |
| add latent-space RL exploration (ablation) | REM · ParisAttack2015, IC, 20% budget | influence spread: 515237.59 -> 637621.59 | Helped |
| add PMoE instead of using one GNN (ablation) | REM · ParisAttack2015, IC, 20% budget | influence spread: 605875.67 -> 637621.59 | Helped |
Surprising findings
- At 10,000 nodes, MIM-Reasoner (7.948s) has lower inference time than REM (8.873s), so REM is not uniformly fastest.
- Under IC at 1% budget on several datasets, ISF and KSN outperform REM (e.g., Cora-ML 398.34 vs. 347.34).
Limitations
- Some baselines (ISF, KSN, GCOMB, ToupleGDD, DeepIM) run out of memory on the larger datasets, so comparisons there are limited.
- Lacks a theoretical guarantee beyond the monotonicity and consistency lemmas, which rely on strong assumptions (converged models, non-decreasing aggregation).
Experimental setup
Datasets: Cora-ML Celegans (BioGRID multiplex GPI) Arabidopsis (BioGRID multiplex) NYClimateMarch2014 Twitter network ParisAttack2015 Twitter network
Benchmarks: Cora-ML Celegans Arabidopsis NYClimateMarch2014 ParisAttack2015
Metrics: total influenced nodes percentage of infected nodes inference wall-clock time
Baselines: ISF KSN GCOMB ToupleGDD DeepIM MIM-Reasoner REM-NonRL REM-NonMixture
Builds on: variational autoencoder noisy top-k gating prioritized experience replay DeepIM MIM-Reasoner
Closest papers by meaning
- Graph Bayesian Optimization for Multiplex Influence Maximization (AAAI 2024 · similarity 0.91)
- DiffIM: Differentiable Influence Minimization with Surrogate Modeling and Continuous Relaxation (AAAI 2025 · similarity 0.87)
- DeepSN: A Sheaf Neural Framework for Influence Maximization (AAAI 2025 · similarity 0.87)
- Hyperparametric Robust and Dynamic Influence Maximization (AAAI 2025 · similarity 0.87)
- Unveiling Environmental Sensitivity of Individual Gains in Influence Maximization (NeurIPS 2025 · similarity 0.86)
- DAG-Aware Variational Autoencoder for Social Propagation Graph Generation (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.