Globalization of local models with SVMs

Local models are high-quality models of small regions of the input space of a learning problem. The advantage of local models is that they are often much more interesting and understandable to the domain expert, as they can concisely describe single aspects of the data instead of describing everything at once as global models do. This results in better plausibility and reliability of models. While it is relatively easy to find an accurate global and interesting local model independently, the combination of both goals is a challenging task, which is hard to solve by current methods. This paper presents an SVM based approach that integrates multiple local classification models into one global model, such that the resulting global model is both accurate and adequately reflects the local patterns on which it is based.