Research map / Label-efficient and robust learning / Recognition and long-tail
Stray Intrusive Outliers-Based Feature Selection on Intra-Class Asymmetric Instance Distribution or Multiple High-Density Clusters
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 SIOFS, a supervised feature selection method that identifies stray intrusive outliers using a refined density-mean center and a skewness-modified 3σ threshold, then scores features by intrusion degree.
Problem
Feature selection for high-dimensional classification where classes have intra-class asymmetric distributions or multiple high-density clusters, making the class body hard to identify and outliers misleading for existing FS methods.
Approach
Focuses on stray intrusive outliers (instances far from their own class body that intrude into another class body), defines a refined density-mean (RDM) center, fuses a skewness coefficient into the 3σ threshold to delimit class bodies, and ranks features by the intrusion degree of these outliers.
Key results
- On the 12 low-level-feature datasets, SIOFS achieves best or second-best ACC on all datasets, e.g., outperforming the second-best by 5.48% on Lung dis at top 50 features and by 4.93% on GISETTE.
- SIOFS achieves best ACC on UCM (94.95%), AID (85.83%) and ModelNet (93.03%) among compared FS methods on deep features.
- On Caltech101 with top 90% features, SIOFS reaches 54.13% ACC versus 53.83% for IOFS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SIOFS instead of the strongest reported baseline | Feature selection and classification · CLL | accuracy (%): 69.22±4.04 -> 71.32±3.84 | Helped |
| Use SIOFS instead of the strongest reported baseline | Feature selection and classification · TOX | accuracy (%): 82.16±6.05 -> 83.24±4.99 | Helped |
| Use SIOFS instead of the strongest reported baseline | Feature selection and classification · Carcinom | accuracy (%): 93.39±1.78 -> 93.77±2.72 | Helped |
| Use SIOFS instead of the strongest reported baseline | Feature selection and classification · Lung | accuracy (%): 94.66±0.77 -> 95.32±1.02 | Helped |
| Use SIOFS instead of the strongest reported baseline | Feature selection and classification · Lung dis | accuracy (%): 86.30±0.00 -> 86.99±1.90 | Helped |
| Use SIOFS instead of the strongest reported baseline | Feature selection and classification · Lymphoma | accuracy (%): 89.24±1.67 -> 89.76±2.20 | Helped |
| Use SIOFS instead of the strongest reported baseline | Feature selection and classification · Nci9 | accuracy (%): 54.72±3.39 -> 54.17±2.10 | Hurt |
| Use SIOFS instead of the strongest reported baseline | Feature selection and classification · GLIOMA | accuracy (%): 75.67±1.80 -> 76.00±2.00 | Helped |
| Use SIOFS instead of the strongest reported baseline | Feature selection and classification · colon | accuracy (%): 82.80±4.02 -> 83.33±1.20 | Helped |
| Use SIOFS instead of the strongest reported baseline | Feature selection and classification · ORL | accuracy (%): 94.33±1.99 -> 94.63±1.64 | Helped |
| Use SIOFS instead of the strongest reported baseline | Feature selection and classification · Yale | accuracy (%): 93.29±2.80 -> 71.21±4.70 | Hurt |
| Use SIOFS instead of the strongest reported baseline | Feature selection and classification · GISETTE | accuracy (%): 71.31±2.26 -> 73.08±6.70 | Helped |
Surprising findings
- Replacing the skewness-based coefficient with a constant 2 (w/o SC) gives nearly identical accuracy on most datasets, with gains concentrated on CLL, TOX and Carcinom.
Limitations
- Runtime is higher on larger datasets such as GISETTE and AID, dominated by the O(n^2) RDM center computation.
- Performance depends on the ratio parameter alpha, which is sensitive for some datasets (e.g., CLL) and requires tuning; the parameter-free variant uses a fixed alpha=0.1.
- Evaluation uses only a linear SVM classifier and ACC/NMI/CM metrics.
Experimental setup
Datasets: CLL TOX Carcinom Lung Lung dis Lymphoma Nci9 GLIOMA colon ORL Yale GISETTE
Benchmarks: CLL TOX Carcinom Lung Lung dis Lymphoma Nci9 GLIOMA colon ORL Yale GISETTE
Metrics: accuracy normalized mutual information confusion matrix
Baselines: Fisher score QMI ReliefF TRC ILFS FSDOC FSTU FSNS ReOLSR MRMSR S2DFS IOFS
Builds on: IOFS FSDOC CFDP 3-sigma principle
Closest papers by meaning
- Out-of-Distribution Detection in Long-Tailed Recognition with Calibrated Outlier Class Learning (AAAI 2024 · similarity 0.84)
- DOS: Diverse Outlier Sampling for Out-of-Distribution Detection (ICLR 2024 · similarity 0.84)
- Mining In-distribution Attributes in Outliers for Out-of-distribution Detection (AAAI 2025 · similarity 0.84)
- How Does Unlabeled Data Provably Help Out-of-Distribution Detection? (ICLR 2024 · similarity 0.83)
- Out-of-Distribution Detection with Prototypical Outlier Proxy (AAAI 2025 · similarity 0.83)
- EAT: Towards Long-Tailed Out-of-Distribution Detection (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.