Research map / Label-efficient and robust learning / Semi-supervised learning
SLRL: Semi-Supervised Local Community Detection Based on Reinforcement Learning
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 SLRL, a reinforcement-learning method for semi-supervised local community detection that extracts a local structure around a given node, selects similar known communities via spectral clustering, and trains an expander with policy gradient to grow the node's community.
Problem
Detecting the community containing a given node using a set of known communities as prior information, where global structure learned from known communities may not fit the target node's community.
Approach
Uses RL-trained extractor to get a rough local community around the node, filters known communities by structural similarity (shortest-path kernel + spectral clustering) to that local structure, and fine-tunes an RL expander (initialized from the extractor) on the selected communities to expand the node's community.
Key results
- On Amazon, SLRL achieves F-score 0.878 vs 0.844 for CommunityAF and 0.828 for SLSS.
- On DBLP, SLRL achieves F-score 0.662 vs 0.626 for SLSS; on Youtube 0.292 vs 0.277 for CommunityAF.
- SLRL underperforms on LiveJournal (F 0.580 vs 0.643 for SLSS), attributed to imbalanced community structure types among known communities.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SLRL instead of the strongest local-detection baseline | SLRL · Amazon, DBLP, and Youtube | Fscore: 0.844 -> 0.878; 0.626 -> 0.662; 0.277 -> 0.292 | Helped |
| use SLRL instead of the strongest local-detection baseline | SLRL · LiveJournal | Fscore: 0.643 -> 0.580 | Hurt |
| use SLRL instead of the strongest local-detection baseline | SLRL · Twitter | Fscore: 0.377 -> 0.378 | Helped |
| use SLRL instead of SEAL or CLARE | SLRL · Amazon, DBLP, LiveJournal, and Twitter | Fscore: 0.839 -> 0.876; 0.625 -> 0.685; 0.679 -> 0.682; 0.285 -> 0.374 | Helped |
| use SLRL instead of SEAL | SLRL · Youtube | Fscore: 0.356 -> 0.361 | Helped |
| add extractor-based local structure extraction (ablation) | SLRL · Amazon and DBLP | Fscore: 0.856 -> 0.873; 0.586 -> 0.613 | Helped |
| add similar-community selection instead of random selection (ablation) | SLRL · Amazon and DBLP | Fscore: 0.852 -> 0.873; 0.577 -> 0.613 | Helped |
| add community expansion (ablation) | SLRL · Amazon and DBLP | Fscore: 0.860 -> 0.873; 0.571 -> 0.613 | Helped |
| vary the discount factor γ from 0.25 to 0.75 (ablation) | SLRL · Amazon, DBLP, Youtube, and Twitter | Fscore | Helped |
| vary the discount factor γ from 0.75 toward 1 (ablation) | SLRL · Amazon | Fscore | Mixed |
| replace spectral clustering with HAC or K-medoids and vary cluster count (ablation) | SLRL · Amazon and DBLP | Fscore: within 0.02 | No clear effect |
Surprising findings
- SLRL is worse than the unsupervised M method and SLSS on LiveJournal, where M performs well, suggesting the method's assumptions mismatch that dataset.
- SLRL underperforms CLARE on Youtube, attributed to CLARE's size restrictions matching actual community sizes.
Limitations
- Requires known communities as prior information.
- Performance drops on LiveJournal due to imbalanced community structure types among known communities.
- Underperforms CLARE on Youtube.
Experimental setup
Datasets: Amazon DBLP LiveJournal Youtube Twitter
Benchmarks: Amazon DBLP LiveJournal Youtube Twitter
Metrics: Precision Recall F-score Jaccard
Baselines: M method SLSS CommunityAF SEAL CLARE
Builds on: SEAL CLARE SLSS iGPN REINFORCE-style policy gradient
Closest papers by meaning
- Semi-supervised Community Detection via Structural Similarity Metrics (ICLR 2023 · similarity 0.88)
- Incorporating Network Embedding into Markov Random Field for Better Community Detection (AAAI 2019 · similarity 0.87)
- Lagrangian Constrained Community Detection (AAAI 2018 · similarity 0.86)
- Learning Persistent Community Structures in Dynamic Networks via Topological Data Analysis (AAAI 2024 · similarity 0.85)
- Community Detection Guarantees using Embeddings Learned by Node2Vec (NeurIPS 2024 · similarity 0.84)
- The Map Equation Goes Neural: Mapping Network Flows with Graph Neural Networks (NeurIPS 2024 · similarity 0.84)
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.