Feature autoencoder

The feature autoencoder is composed of an encoder and decoder process. It takes the gene expression matrix as input, and LTMG performs regularization processing to learn the embedding expression of scRNA-seq data. The encoder constructs the embedding of low dimension X^ (reconstructed gene expression matrix) through the input gene expression matrix X (normalized gene expression matrix), then the decoder is reconstructed according to embedding, so the encoder is a process of dimension reduction. The feature autoencoder is trained by minimizing the loss function of the difference between the input gene expression matrix and the output matrix so that the output matrix is as similar as possible to the input matrix. The mean square error (MSE) is defined as: