부산대학교 도서관

In this paper, the classification problem of schizophrenia patients from healthy controls is considered, whose goal is to explore the relationship between DNA characteristics and schizophrenia. However, the DNA methylation data has the properties of small samples in high dimension and non-Gaussian distribution which makes it hard to do classification with DNA methylation data. Hence a classification method based on deep learning is designed. We propose a feature selection method based on attention mechanism which embeds a weight gated layer in the network structure to get a task-related sparse representation of the DNA methylation data. The performance of proposed method outperforms existing feature selection methods. On a real-world data set, the classification with proposed method achieves a high accuracy.