The output coding for multiclass learning problems is a generalization of one-per-class, all-pairs, and error correcting output codes. Although, the prevailing concepts of output coding has been error correcting properties, the one-per-class and all-pairs are still considered to be one of the state-of-art methods. However, these two methods are contrary to each other in the aspect of producing complex dichotomies and the problem of nonsense outputs. In additions, they all perform a prior decomposition without regards to the properties of a given training data set. In this paper, we propose a new data-driven output coding method that is the generalized form of one-per-class and all-pairs. We present the properties of the proposed method. From experimental results on both a toy problem and real benchmark datasets, we present that our proposed method achieves a comparable performance with good properties.