NobSVM-RFE: An Improved SVM-RFE Algorithm for Feature Selection

With the popularization of DNA microarray technology, gene data classification has become one of the major frontiers in bioinformatics. The complexity of mi-croarray experiments leads to gene expression profiling data with small samples and high dimensionality. Therefore, this paper proposes a new feature selection algorithm as unbiased support vector machine recursive feature elimination (NobSVM-RFE), which improves the training method of SVM-RFE and uses unbiased support vector machine to train the original feature set. NobSVM-RFE calculates the score of each feature and recursively eliminates features with the lowest score from the original feature set until the number of features in the origi-nal feature set is reduced to a pre-determined value. Experimental results on three publicly datasets show that after NobSVM-RFE preprocessing, SVM obtained 98.45%, 97.14% and 98.10% accuracy and 96.25%, 97.54% and 97.77% F1-score respectively. Meanwhile, the performance of the enhanced SVM-RFE is significantly improved compared to the traditional feature selection algorithm, which demonstrates the effectiveness of the NobSVM-RFE algorithm.