New framework enhances feature selection for complex AI models

Researchers have developed a new framework for controlling False Discovery Rate (FDR) in feature selection, specifically addressing challenges in sequential and grouped models. This method extends beyond traditional coordinate-wise hypothesis testing to handle scenarios where a single original feature is represented by multiple sub-features, such as in recurrent neural networks or attention mechanisms. The proposed approach utilizes grouped-feature FDR control, constructing null-symmetric mirror statistics for linear models and combining Permutation SHAP derivatives with kernel-based dependence measures for neural sequential models. This model-agnostic framework does not require assumptions about covariate distribution and has demonstrated reliable FDR control and improved power in experiments. AI

IMPACT This research could improve the reliability and power of feature selection in complex AI models, leading to more accurate and interpretable results.

RANK_REASON The item is an academic paper detailing a new statistical framework for machine learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Franklin Delano Roosevelt
• Gaussian Mirror
• Gotit.pub
• Grouped Linear Models
• Group Gaussian Mirror
• Hugging Face
• Influence Flower
• Neural Gaussian Mirror
• Neural Sequential Models
• Permutation SHAP
• ScienceCast

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →