To improve classific precision of network intrusion detection model and reduce the number of training data set and learning time, a new supervisal algorithm based on ε Support Vector Regression machines (ε-SVR) machine was proposed. Firstly, normalization was used on training data set, and then a new coefficience of sparse penalty function was adjusted. Finally, the experimental results using KDD CUP 1999 data set show that this approach can detect intrusion behavior, increase its veracity and validity, and reduce its distortion.