New IF-dRVFL models boost noisy data classification accuracy
Researchers propose IF-dRVFL and IF-edRVFL to handle noisy data better than current SOTA models.
Researchers from India and Singapore have developed two novel models—IF-dRVFL and IF-edRVFL—to address the limitations of current state-of-the-art deep randomized neural networks like dRVFL and edRVFL. These new models leverage intuitionistic fuzzy theory to assign adaptive weights to training samples, effectively distinguishing between clean, noisy, and outlier data points.
The proposed frameworks compute membership and non-membership degrees in the kernel space, using distance metrics from class centroids and local neighborhood heterogeneity. In extensive experiments on UCI and KEEL benchmark datasets—including scenarios with added Gaussian noise—the IF-dRVFL and IF-edRVFL models demonstrated superior performance over existing fuzzy and non-fuzzy approaches, highlighting their robustness in real-world noisy datasets.
- Introduces IF-dRVFL and IF-edRVFL, uncertainty-aware deep randomized neural networks designed to handle noisy and outlier data
- Achieves superior performance on UCI and KEEL benchmarks with Gaussian noise, outperforming existing SOTA models
- Source code is publicly available for implementation and further research
Why It Matters
Enables more reliable AI models in noisy real-world environments, improving accuracy in critical applications.