Expectation Propagation of Gaussian Process Classification and Its Application to Gene Expression Analysis

Expectation Propagation (EP) is an approximate Bayesian inference technique which has been applied to Gaussian Process Classification (GPC) [4]. In this paper, we investigate four different likelihood functions of GPC, and present EP algorithms for each of these four models. We compare the performances of these models on synthetic data in circular shape. Comparative study is performed on EP-GPC with SVM and Laplace-GPC. Experimental results show EP-GPC outperforms the other two kernel methods on high dimensional gene expression data. A feature selection technique, Automatic Relevance Determination (ARD), is applied to find the relevance of genes. Experiments show the effectiveness of ARD for all classification models.