Correlation coefficient between the expression levels of two genes plays an important role in the inference of their relationship in microarray experiments. Gene expression data before normalization often present high correlation coefficients among a large proportion of genes. Some of these high correlations are caused by changes in gene expression levels. However, most of them are caused by systematic errors. It is intended to eliminate superficial high correlations induced by systematic errors and at the same time, preserve high correlation coefficients stem from gene interactions. Although there are a number of comparisons among different normalization methods, less work focused on evaluating the effect of normalization procedures on correlation coefficients among genes and which method does the best in restoring gene correlation structure. Some gene expression data were simulated with reference to real world gene expression data. With the help of these simulated data, it was determined which normalization method does the best in restoring gene correlation structure. In addition, it was shown that the simulated data and the real world data have the same gene correlation structure, so the conclusion drawn from simulated data can be applied to the real world. For 5 normalization methods compared here, it can be concluded that the loess method is the most appropriate one in eliminating superficial correlation coefficients.