Feature selection is a necessary processing step for class prediction using microarray expression data. Traditional methods select top-ranked genes in terms of their discriminative powers. This strategy unavoidably results in redundancy, whereby correlated features with comparable discriminative powers are equally favorable. Redundancy has many drawbacks among other aspects. As a result, reducing redundancy is an important goal for most feature selection methods. Almost all methods for redundancy reduction are based on the correlation between gene expression levels. In this paper, we utilize the knowledge in gene ontology to provide a new model for measuring redundancy among genes. We propose a novel method to evaluate the GO semantic similarity and a similarity metric, which incorporates semantic and expression level similarities. We compare our method with traditional expression value-only similarity model on several public microarray datasets. The experimental results show that our approach is capable of offering higher or the same classification accuracy while providing a smaller gene feature subset.