Gene expression profiles play more and more important roles in accurate tumor diagnosis and treatment. However, the curse of dimensionality that the number of genes far exceeds the number of samples issues the challenges to the traditional dimensionality reduction methods. Here based on two-stage dimensionality reduction model we design 18 tumor classification methods by combining two classical gene filters with three common dimensionality reduction methods: principal component analysis (PCA), linear discriminative analysis (LDA) and multidimensional scaling (MDS) method to extract discriminative features and use three common machine learning methods to evaluate the prediction accuracy of the extracted features on six tumor datasets, respectively. Although gene expression presents the non-linear characteristics, non-linear dimensionality reduction method MDS is not always the best in prediction accuracy among the three dimensionality reductions on all six tumor datasets. Moreover, the performance comparison indicates that no single dimensionality reduction is always superior to the others on all of the six tumor datasets. Our results also suggest that the prediction accuracy obtained depends strongly on the dataset, and less on the gene selection and classification methods.