Reproduce “Open Set Recognition for Random Forest” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Guanchao Feng, Dhruv Desai, Stefano Pasquali, Dhagash Mehta
Researchers at BlackRock, Inc. developed RF-KOSNN, an approach that enables Random Forest classifiers to perform open-set recognition by integrating RF-GAP proximities with Gaussian process-based distance metric learning and a global Extreme Value Theory model for unknown class detection. RF-KOSNN consistently outperformed existing open-set recognition baselines across multiple datasets, including synthetic, image, and financial data, by accurately classifying known classes and effectively identifying unknown samples.
Key insights
- Random Forest proximity measures (RF-GAP) can be effectively combined with distance metric learning to create a feature space optimized for open-set recognition.
- Gaussian processes provide a robust mechanism for learning a Mahalanobis distance metric from RF-GAP proximities, enhancing separation between known and unknown classes.
- A global Extreme Value Theory model for unknown detection proves more robust and scalable than class-wise models, particularly in scenarios with numerous classes or sparse data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.