Error-correcting output codes (ECOCs) have been widely used in many applications for multi-class classification problems. The problem is that ECOCs cannot be ap- plied directly on two-class datasets. The goal of this thesis is to design and evaluate an approach to solve this problem, and then investigate whether the approach can yield better classification models. To be able to use ECOCs, we turn two-class datasets into multi-class datasets first, by using clustering. With the resulting multi-class datasets in hand, we evalu- ate three different encoding methods for ECOCs: exhaustive coding, random coding and a “pre-defined” code that is found using random search. The exhaustive coding method has the highest error-correcting abilities. However, this method is limited due to the exponential growth of bit columns in the codeword matrix precluding it from being used for problems with large numbers of classes. Random coding can be used to cover situations with large numbers of classes in the data. To improve on completely random matrices, “pre-defined” codeword matrices can be generated by using random search that optimizes row separation yielding better error correction than a purely random matrix. To speed up the process of finding good matrices, GPU parallel programming is investigated in this thesis. From the empirical results, we can say that the new algorithm, which applies multi-class ECOCs on two-class data using clustering, does improve the performance for some base learners, when compared to applying them directly to the original two- class datasets.