A decision directed acyclic graph support vector machine is a typical multi-class classification with support vector machines. But error accumulation exists in the traditional decision classification, and its generalization ability depends on the tree structure. In this paper, to improve the generalization ability of DDAGSVM, a novel separable measure is defined based on the generalized KKT, and an improved decision directed acyclic graph support vector machine is given. The three-spiral and HRRP experimental results show that this kind of algorithm has an obvious effectiveness in controlling classification errors.