CatNet: Controlling False Discovery Rate in LSTM with SHAP and Gaussian Mirrors
This academic paper introduces CatNet, a novel algorithm designed to control the False Discovery Rate (FDR) and select significant features within Long Short-Term Memory (LSTM) networks. The method utilizes the derivative of SHAP values to quantify feature importance and constructs a vector-formed mirror statistic using the Gaussian Mirror algorithm for effective FDR control. To address instability arising from nonlinear or temporal correlations among features, the authors propose a new kernel-based independence measure. The study demonstrates that CatNet performs robustly across various model settings using both simulated and real-world data, effectively reducing overfitting and enhancing model interpretability. Although the framework shows promise for extension to other time-series or sequential deep learning models, the specific version of this paper (v4) on arXiv has been withdrawn by the author, Jiaan Han. The research contributes to the fields of machine learning and statistics by integrating explainable AI techniques with statistical rigor in deep learning applications.
Wire timeline
CatNet: Controlling False Discovery Rate in LSTM with SHAP and Gaussian Mirrors
This academic paper introduces CatNet, a novel algorithm designed to control the False Discovery Rate (FDR) and select significant features within Long Short-Term Memory (LSTM) networks. The method utilizes the derivative of SHAP values to quantify feature importance and constructs a vector-formed mirror statistic using the Gaussian Mirror algorithm for effective FDR control. To address instability arising from nonlinear or temporal correlations among features, the authors propose a new kernel-based independence measure. The study demonstrates that CatNet performs robustly across various model settings using both simulated and real-world data, effectively reducing overfitting and enhancing model interpretability. Although the framework shows promise for extension to other time-series or sequential deep learning models, the specific version of this paper (v4) on arXiv has been withdrawn by the author, Jiaan Han. The research contributes to the fields of machine learning and statistics by integrating explainable AI techniques with statistical rigor in deep learning applications.
cs.AI updates on arXiv.org