Research map / Label-efficient and robust learning / Noisy labels
KFNN: K-Free Nearest Neighbor For Crowdsourcing
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 KFNN, a label integration algorithm for crowdsourcing that automatically determines each instance's neighborhood size via Mahalanobis-distance label enhancement, Kalman-filtered class margins, and max-margin selection.
Problem
Label integration with only a few noisy labels per instance in crowdsourcing, where existing neighbor-based methods use a fixed K for all instances, which is unrealistic.
Approach
Replaces the fixed KNN neighborhood with a per-instance neighborhood size chosen as the argmax of a Kalman-filtered class margin over a neighbor sequence ordered by Mahalanobis-based distance, with the label distribution enhanced by a distance-derived potential label distribution.
Key results
- KFNN achieves average Macro-F1 of 79.64% on 34 simulated datasets, versus 76.68% for the next best method MNLDP.
- On Income and Leaves real-world datasets, KFNN achieves the highest Macro-F1 (77.17% and 65.40%) among compared methods.
- Wilcoxon tests show KFNN significantly outperforms all six comparison algorithms in Macro-F1 at alpha=0.05.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use KFNN instead of MNLDP | label integration · 34 simulated datasets (average) | Macro-F1: 76.68 -> 79.64 | Helped |
| use KFNN instead of LAWMV | label integration · 34 simulated datasets (average) | integration accuracy: 87.72 -> 86.24 | Hurt |
| use KFNN instead of LAGNN | label integration · Income | Macro-F1: 76.42 -> 78.69 | Helped |
| use KFNN instead of LAWMV | label integration · Income | integration accuracy: 74.00 -> 77.17 | Helped |
| use KFNN instead of AALI | label integration · Leaves | Macro-F1: 60.20 -> 65.40 | Helped |
| use KFNN instead of AALI | label integration · Leaves | integration accuracy: 67.19 -> 72.14 | Helped |
| add K-free optimization (KF) (ablation) | KFNN on Income · Income | Macro-F1: 77.95 -> 78.69 | Helped |
| add label distribution enhancement (LDE) (ablation) | KFNN on Income · Income | Macro-F1: 76.13 -> 78.69 | Helped |
| add K-free optimization (KF) and label distribution enhancement (LDE) (ablation) | KFNN on Income · Income | Macro-F1: 75.22 -> 78.69 | Helped |
| increase Kalman filter parameter β toward 1 | KFNN · Income and Leaves | Macro-F1 | Helped |
| vary Kalman filter parameter α | KFNN · Income and Leaves | Macro-F1 | No clear effect |
Surprising findings
- In integration accuracy on simulated data, KFNN (86.24%) is slightly below LAWMV (87.72%) and comparable to MNLDP, so it does not dominate on accuracy.
- The Mahalanobis distance metric is not effective across all datasets (e.g., autos and breast-cancer), as acknowledged by the authors.
Limitations
- Kalman filter parameters alpha and beta are fixed empirically and do not adapt to datasets.
- Max-min normalization transforming distances into a potential label distribution is coarse.
- Distance metric is not effective on some datasets, degrading performance.
Experimental setup
Datasets: 34 simulated datasets from UCI-style collections via CEKA/WEKA Income (AMT) Leaves (AMT)
Benchmarks: 34 simulated crowdsourced datasets Income Leaves
Metrics: Macro-F1 integration accuracy
Baselines: MV IWMV AALI LAGNN LAWMV MNLDP
Builds on: LAWMV MNLDP label enhancement majority voting
Closest papers by meaning
- Towards a Foundation Model for Crowdsourced Label Aggregation (ICLR 2026 · similarity 0.87)
- Semi-Supervised Learning From Crowds Using Deep Generative Models (AAAI 2018 · similarity 0.86)
- TLLC: Transfer Learning-based Label Completion for Crowdsourcing (ICML 2025 · similarity 0.85)
- Label Distribution Propagation-based Label Completion for Crowdsourcing (ICML 2025 · similarity 0.85)
- IWBVT: Instance Weighting-based Bias-Variance Trade-off for Crowdsourcing (NeurIPS 2024 · similarity 0.84)
- Coupled Confusion Correction: Learning from Crowds with Sparse Annotations (AAAI 2024 · similarity 0.83)
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.