CatNet: Controlling the False Discovery Rate in LSTM with SHAP Feature Importance and Gaussian Mirrors

Researchers introduce CatNet, an algorithm that controls False Discovery Rate (FDR) in LSTM neural networks by combining SHAP feature importance derivatives with a Gaussian Mirror statistical approach. The method addresses overfitting and model interpretability challenges in time-series deep learning through improved feature selection and a novel kernel-based independence measure.

CatNet represents a meaningful advancement in making deep learning models more interpretable and statistically rigorous, particularly for time-series applications. The algorithm tackles a fundamental challenge in machine learning: distinguishing genuinely important features from spurious correlations that can degrade model performance. By integrating SHAP (SHapley Additive exPlanations) values with formal FDR control methods, the researchers create a framework that balances predictive accuracy with statistical rigor.

The innovation stems from longstanding tensions between model complexity and interpretability. Traditional LSTM networks excel at capturing temporal patterns but often become black boxes that overfit to noise. Prior FDR control methods struggled with neural networks due to their nonlinearity and feature interdependencies. CatNet's novel kernel-based independence measure directly addresses this limitation, enabling proper feature selection even when inputs exhibit temporal or nonlinear correlations.

For practitioners developing financial prediction models, medical forecasting systems, or sensor-based applications, this work offers practical value. Better feature selection reduces computational overhead, accelerates model training, and produces more reliable predictions on unseen data. The framework's extensibility to other sequential deep learning architectures amplifies its potential impact across diverse domains.

The significance lies not in revolutionary performance gains but in methodological rigor. Regulators and institutions increasingly demand explainable AI, making techniques that provide statistical guarantees on feature importance valuable. As deep learning adoption expands in regulated industries, tools that formally control false discovery rates help bridge the gap between model sophistication and accountability requirements.

• →CatNet combines SHAP feature importance with Gaussian Mirror FDR control to identify statistically significant LSTM features
• →Novel kernel-based independence measure handles nonlinear and temporal correlations previously problematic for FDR algorithms
• →Framework reduces overfitting and improves model interpretability on both simulated and real-world datasets
• →Method extends beyond LSTM to other time-series and sequential deep learning architectures
• →Addresses growing demand for explainable AI with formal statistical guarantees on feature selection