RoBell-RVFL: A Robust Generalized Bell Random Vector Functional Link Network
RoBell-RVFL is a robust, lightweight randomized neural network that addresses class imbalance and label noise through a quality-aware, sample-level weighting mechanism It uses a dual-strategy approach: unit weights for minority class samples to preserve their information, and a probability-weighted generalized bell (gbell) membership function in kernel-induced feature space for majority class samples The method suppresses noisy, boundary, and outlier samples within the majority class, enabling l
Analysis
TL;DR
- RoBell-RVFL is a robust, lightweight randomized neural network that addresses class imbalance and label noise through a quality-aware, sample-level weighting mechanism
- It uses a dual-strategy approach: unit weights for minority class samples to preserve their information, and a probability-weighted generalized bell (gbell) membership function in kernel-induced feature space for majority class samples
- The method suppresses noisy, boundary, and outlier samples within the majority class, enabling learning from informative rather than merely abundant samples
- Extensive evaluations on UCI and KEEL benchmarks show consistent and significant outperformance over state-of-the-art RVFL variants, with robustness maintained under up to 40% label noise
- The work demonstrates that adaptive, quality-aware sample weighting is essential for robust RVFL learning, rendering conventional global weighting schemes ineffective in noisy and imbalanced environments
Why It Matters
This research addresses a critical gap in randomized neural networks by moving beyond simple class-proportion adjustments to incorporate intra-class distribution and sample quality, which is essential for real-world deployment where data is rarely clean or balanced. For AI practitioners working with imbalanced datasets—particularly in domains like fraud detection, medical diagnosis, or anomaly detection—RoBell-RVFL offers a computationally efficient alternative to expensive resampling or deep learning approaches while maintaining closed-form learning speed.
Technical Details
- Architecture: Builds on Random Vector Functional Link (RVFL) networks, a class of single-hidden-layer feedforward networks with randomized hidden nodes and closed-form output weight learning
- Dual-Strategy Weighting: Minority class samples are assigned unit weights to strictly preserve their information; majority class samples are adaptively regulated using a probability-weighted generalized bell (gbell) membership function operating in a kernel-induced feature space
- Quality-Aware Mechanism: Explicitly incorporates local class probability and class distribution information into the learning process, enabling adaptive control over sample contributions based on their informational value rather than mere abundance
- Robustness Testing: Evaluated under up to 40% label noise, demonstrating resilience that conventional global weighting schemes fail to achieve
- Benchmarks: Tested on UCI and KEEL benchmark datasets, outperforming recent state-of-the-art RVFL variants
Industry Insight
- The shift from global to adaptive, quality-aware sample weighting represents a paradigm shift for imbalanced learning in lightweight neural architectures, suggesting that future RVFL variants should prioritize intra-class distribution modeling over simple oversampling
- For practitioners dealing with noisy, imbalanced real-world data, RoBell-RVFL offers a compelling trade-off between computational efficiency and robustness, making it suitable for resource-constrained or latency-sensitive applications where deep learning is impractical
- The demonstrated resilience to 40% label noise highlights the importance of sample-level quality assessment in training pipelines, encouraging the integration of similar weighting mechanisms into broader randomized network frameworks and industrial ML workflows
Disclaimer: The above content is generated by AI and is for reference only.