Partitioning trees: A global multiclass classification technique for SVMs

Presented in this paper is a novel technique for multiclass classification in SVMs through combination of binary classifiers,namely that of Partitioning Trees (P-Trees). The technique aims at improving the Directed Acyclic Graphs (DAGs) both interms of training as well as testing performance. It works by progressively constructing a decision graph, where each node is abinary classifier. Each trained node defines a dichotomy over the instance space which, in turn, is used to train subsequent nodes.In this way, every node trains against only a subset of the samples of its classes; namely the samples that reach the node throughthe decision graph in addition to a subsampled version of the ones that fail to reach it. Training sets reduce in size and decisionsurfaces become more compact, thus improving training and testing performance. Extensive experimental results demonstratethe effectiveness of the proposed technique in reducing the training and testing time in SVMs, while maintaining comparablegeneralization performance to the 1vs1 and DAGs techniques.