Current support vector machines (SVMs) for large-scale datasets classification problems are almost sensitive to noises. To overcome this problem

a new soft kernel convex hull support vector machine called SCH-SVM is proposed based on the soft kernel convex hull and pinball loss function. SCH-SVM extracts the soft convex hull vectors in the kernel space

which can represent geometric profile of data in the kernel space. Then SCH-SVM represents the original samples which projected as the soft convex hull vectors for the training samples

and finds the maximum quantile distance between soft kernel convex hulls belonging to two classes by using pinball loss function. Theoretical analysis and numerical experiments show that SCH-SVM has distinctive ability of training time

noise resistibility

and the number of support vectors.