Research map / Graph learning / Graph self-supervised learning
Globally Correlation-Aware Hard Negative Generation
Wenjie Peng, Hongxiang Huang, Tianshui Chen, Qiuping Ke, Gang Dai, Shuangping Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GCA-HNG hard negative generation to N-pair+ loss | ResNet-50 R512 · CUB-200-2011 | Recall@1: 66.5 -> 70.3 | Improved |
| add GCA-HNG to Proxy Anchor loss | ResNet-50 R512 · Cars196 | Recall@1: 88.4 -> 91.1 | Improved |
| add GCA-HNG to Proxy Anchor loss | DINO-S DN384 · SOP | Recall@1: 85.2 -> 85.9 | Improved |
| add global sample correlations (node message propagation) (ablation) | GCA-HNG NP+ · CUB-200-2011 | Recall@1: 69.1 -> 70.3 | Improved |
| add global sample correlations (node message propagation) (ablation) | GCA-HNG PA · CUB-200-2011 | Recall@1: 69.3 -> 72.3 | Improved |
| add Hadamard product addition of edge info in node propagation (ablation) | GCA-HNG PA · CUB-200-2011 | Recall@1: 71.7 -> 72.3 | Improved |
| use channel-wise random-weighted interpolation instead of single coefficient (ablation) | GCA-HNG NP+ · CUB-200-2011 | Recall@1: 68.0 -> 70.3 | Improved |
| combine interpolated representations by random weighting (ablation) | GCA-HNG NP+ · CUB-200-2011 | Recall@1: 69.9 -> 70.3 | Improved |
| add cross-entropy loss J_ce to generation loss (ablation) | GCA-HNG NP+ · CUB-200-2011 | Recall@1: 67.7 -> 70.3 | Improved |
| add similarity loss J_sim and diversity loss J_div (ablation) | GCA-HNG PA · CUB-200-2011 | Recall@1: 71.3 -> 72.3 | Improved |
| add GNN loss without hard negative generation (ablation) | Baseline NP+ · CUB-200-2011 | Recall@1: 66.5 -> 66.8 | No clear effect |
| add GNN module (training-time only) | GCA-HNG NP+ · Training time | sec/iter: 0.22 sec/iter -> 0.29 sec/iter | Hurt |
Closest papers by meaning
- Deep Global-sense Hard-negative Discriminative Generation Hashing for Cross-modal Retrieval (ICLR 2026 · similarity 0.86)
- Adaptive Hardness Negative Sampling for Collaborative Filtering (AAAI 2024 · similarity 0.86)
- Does Negative Sampling Matter? a Review With Insights Into its Theory and Applications (TPAMI 2024 · similarity 0.85)
- Contrastive Learning with Hard Negative Samples (ICLR 2021 · similarity 0.85)
- Mathematical Justification of Hard Negative Mining via Isometric Approximation Theorem (ICLR 2024 · similarity 0.84)
- Does GCL Need a Large Number of Negative Samples? Enhancing Graph Contrastive Learning with Effective and Efficient Negative Sampling (AAAI 2025 · 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.
