| Multiclass classification is an important machine learning problem that involves classifying a pattern into one of several classes, and encompasses domains such as handwritten character recognition, protein structure classification, heartbeat arrhythmia identification and many others. In this thesis, we investigate three issues in combining models to perform multiclass classification. First, we demonstrate that ridge regularization is essential in linear combinations of multiclass classifiers. Second, we show that when solving a multiclass problem using a combination of binary classifiers, it is more effective to share hyperparameters across models than to optimize them independently. Third, we introduce a new method for combining binary pairwise classifiers that overcomes several problems with existing pairwise classification schemes and exhibits significantly better performance on many problems. Our contributions span the themes of model selection and reduction from multiclass to binary classification. |
