Improving the one-against-all binary approach for multiclass classification using balancing techniques

摘要

One-against-one and one-against-all are common approaches to break down multiclass classification problems into binary classification problems and build a multiclass classifier. The former approach often yields better multiclass classifiers than the latter due to its structure. The one-against-all approach strengthens or sometimes creates linear inseparability and class imbalance in the binary classifiers during the training phase. In this sense, balancing techniques can be applied to handle the binary imbalance problem and motivate the use of the computationally simpler approach. The one-against-all approach with balancing techniques proposed in this work reaches better accuracy values than the pure one-against-all approach for 7 out of 8 datasets and shows a considerable increase in the weighted recall value for 4 out of 8 datasets. Besides, the accuracy values of the one-against-all approach with balancing techniques are considerably closer to the ones found by the one-against-one approach with less computational efforts.