SVAD: Stacked Variational Autoencoder Deep Neural Network-Based Dimensionality Reduction and Classification of Small Sample Size and High Dimensional Data

The “curse of dimensionality” is a major concern in the field of computational biology, especially when there are many fewer samples then to the number of features. Many strategies for dimension reduction have been developed; however, they all have drawbacks when it comes to high-dimensional and small sample size data, such as high value of variance gradients and over-fitting.To address these issues, we proposed a dynamic variational autoencoder based deep neural network architecture, based on a mathematical foundation for unsupervised learning. The objective of this research is to propose a low-error classification algorithm for limited sample numbers and high-dimensional datasets. The study's innovation is that it guarantees the permissible dimension size regardless of reduction, in contrast to several previous approaches that typically reduce the dimension too heavily. The experimental findings reveal that the suggested method outperforms existing traditional methods such as RNN, CNN, and deep network architecture.