For the tumor gene expression profile data that aiming to high-dimension small samples, how to select the classification feature of samples among thousands genes effectively is the difficult problems for analysis on tumor gene expression profile. First to partition the data set into K average divisions, to use Lasso method performing feature selection on each respectively, and then merge each selected division of subset together to perform feather selection again, and get the final feature gene. This experiment adopts the Support Vector Machine (SVM) as classifier, to take the classification performance of feature gene set by Leave One Out Cross-Validation (LOOCV) method as evaluation standard, improve classification accuracy and with algorithm in good stability. Because of lowered dimensions in each time of calculation, it solves the problem of overhead computational-expensive, and also solves the problem of “over-fitting” in a certain grade. Thus it gets conclusion that the K-partitioning Lasso method shall be an effective method for tumor feature gene selection.