Distribution-dependent feature selection for deep neural networks

While deep neural networks (DNNs) have achieved impressive performance on a wide variety of tasks, the black-box nature hinders their applicability to high-risk, decision-making fields. In such fields, besides accurate prediction, it is also desired to provide interpretable insights into DNNs, e.g., screening important features based on their contributions to predictive accuracy. To improve the interpretability of DNNs, this paper originally proposes a new feature selection algorithm for DNNs by integrating the knockoff technique and the distribution information of irrelevant features. With the help of knockoff features and central limit theorem, we state that the irrelevant feature's statistic follows a known Gaussian distribution under mild conditions. This information is applied in hypothesis testing to discover key features associated with the DNNs. Empirical evaluations on simulated data demonstrate that the proposed method can select more true informative features with higher F-1 scores. Meanwhile, the Friedman test and the post-hoc Nemenyi test are employed to validate the superiority of the proposed method. Then we apply our method to Coronal Mass Ejections (CME) data and uncover the key features which contribute to the DNN-based CME arrival time.