In the disease diagnosis approach of combining with Gene Algorithm (GA) and Support Vector Machine (SVM) ensemble, the attribute redundancy problem still exists. A decision method for diagnosis of breast cancer was proposed based on attribute reduction and SVM. The proposed attribute reduction method took minimizing the attribute number, maximizing the difference attribute number in discernibility matrix and maximizing the dependency degree of condition reduction attributes on decision attributes as the fitness function of GA. After attribute reduction, multiple attribute subsets were selected for SVM ensemble learning. Compared with SVM, experimental results on the breast cancer dataset from UCI databases validate that the classification accuracy increases by 2 percent at least.