Reproduce “CatNet: Controlling the False Discovery Rate in LSTM with SHAP Feature Importance and Gaussian Mirrors” on OpenResearch

About this paper

Jiaan Han, Junxiao Chen, Yanzhe Fu

CatNet offers a novel algorithm to control the False Discovery Rate and identify truly significant features within LSTM networks, particularly for time-series data. It achieves this by combining SHAP feature importance with an extended Gaussian Mirror approach that accounts for non-linear and temporal cross-correlations. On S&P 500 stock prediction, CatNet-selected features reduced NRMSE to 0.042 and increased R^2 to 0.953, demonstrating enhanced predictive performance and model interpretability.

Key insights

• The derivative of SHAP values offers a statistically sound and interaction-aware metric for dynamic feature importance, overcoming the orthogonality assumptions of prior derivative-based methods.
• Explicitly modeling and minimizing time-series cross-correlation through kernel-based dependence measures is vital for constructing stable mirror variables, which in turn enhances the robustness of FDR control in sequential deep learning models.
• Applying rigorous statistical FDR control to feature selection in deep learning can lead to more interpretable models that make fewer false discoveries, while simultaneously boosting predictive accuracy in real-world tasks.

Summary by alphaXiv.