Overcoming the Curse of Dimensionality in Microarray Data Via Generative Data Augmentation and L1-Regularized Selection

Abstract

Feature selection research is crucial to overcome the dimensionality curse. Despite numerous attempts to select features for high-dimensional data, many existing techniques suffer from limited computational efficiency and fail to capture complex and deep correlations among features, particularly in highly complex and low-sample-size. This article presents an innovative framework that uses conditional competitive generative networks (CGANs) to mitigate the limitations of differential evolution (DE)

and the low-dimensional, small-scale dataset problem (HDLSS) by increasing dataset size and improving the stability of the selection process, thereby enhancing the model's generalizability. Experimental results show that the model achieved a classification accuracy of 89.59%, surpassing the best reference method of 86.10%. This highlights the framework's effectiveness in balancing accuracy and computational efficiency, as well as its applicability to diverse high-dimensional data scenarios.