Enhanced Outlier Detection in Microarray Data: A Variational Autoencoder Approach with Bayesian Optimization

The increasing integration of deep neural network methods in recent research has significantly contributed to the growing popularity of outlier detection tasks, as these advanced techniques offer enhanced capabilities for identifying anomalies in complex and high-dimensional datasets. Various techniques, such as clustering-based, distance-based, and density-based methods for outlier detection, have been introduced in recent years. Variational Autoencoders (VAEs) have emerged as a powerful tool for outlier detection, leveraging their capacity to model complex data distributions. However, VAEs exhibit a certain level of complexity, necessitating meticulous adjustment of hyperparameters. This complexity raises the risk of overfitting, particularly in the context of those datasets dealing with high dimensionality, which can result in unreliable outlier detection outcomes. In the proposed technique after special steps of preprocessing, we define a VAE architecture and then set up a Bayesian optimization to tackle with the problem of adjusting hyperparameters. After running the Bayesian optimization and we evaluate our model on outlier detection in two benchmark microarray datasets including Leukemia and DLBCL from Gene Expression Omnibus database. In conclusion, our technique demonstrates superior performance in outlier detection, achieving an exceptional F۱ score that outperforms several recent studies in the field.

Keywords:

outlier detection , variational autoencoder , bayesian optimization , unsupervised , microarray