DiffKnock: Diffusion-based Knockoff Statistics for Neural Networks Inference

2025-10-01

AI Summary

Existing knockoff methods for high-dimensional feature selection struggle to model complex dependencies and nonlinear associations among features, while also lacking finite-sample false discovery rate (FDR) control. To address these limitations, this paper introduces DiffKnockoff—the first framework to integrate diffusion models into knockoff variable construction. DiffKnockoff leverages the iterative denoising process of diffusion models to accurately capture higher-order dependencies in the original features. It further designs an antisymmetric importance statistic by combining neural network gradients with filtered test statistics, enabling high statistical power for detecting nonlinear associations while guaranteeing rigorous finite-sample FDR control. Experiments demonstrate that DiffKnockoff significantly outperforms autoencoder-based baselines on synthetic data. In application to single-cell RNA-seq analysis, it successfully identifies key regulatory genes in the NF-κB pathway, confirming its biological interpretability and practical utility.

Technology Category

• Machine Learning: Dimensionality Reduction/Feature Selection
• Computer Vision: Diffusion Models for Vision
• Data Mining & Knowledge Management: Anomaly/Outlier Detection

Application Category

• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

We introduce DiffKnock, a diffusion-based knockoff framework for high-dimensional feature selection with finite-sample false discovery rate (FDR) control. DiffKnock addresses two key limitations of existing knockoff methods: preserving complex feature dependencies and detecting non-linear associations. Our approach trains diffusion models to generate valid knockoffs and uses neural network--based gradient and filter statistics to construct antisymmetric feature importance measures. Through simulations, we showed that DiffKnock achieved higher power than autoencoder-based knockoffs while maintaining target FDR, indicating its superior performance in scenarios involving complex non-linear architectures. Applied to murine single-cell RNA-seq data of LPS-stimulated macrophages, DiffKnock identifies canonical NF-$κ$B target genes (Ccl3, Hmox1) and regulators (Fosb, Pdgfb). These results highlight that, by combining the flexibility of deep generative models with rigorous statistical guarantees, DiffKnock is a powerful and reliable tool for analyzing single-cell RNA-seq data, as well as high-dimensional and structured data in other domains.

Problem

• Controls false discovery rate in high-dimensional feature selection
• Preserves complex dependencies and detects nonlinear associations
• Generates valid knockoffs using diffusion models for neural networks

Innovation

• Uses diffusion models to generate valid knockoffs
• Employs neural network-based statistics for feature importance
• Combines deep generative models with statistical guarantees