Customization for classification

Abstract

Customization is a machine learning approach in modifying a learned model to be better adapted with additional data. In contrast to training a new model from the additional data, customization relies on the hypothesis that knowing the learned model will serve as useful information which could improve the result of machine learning. However, a common limitation for customization algorithms is that they are specific on the type of model. This results in requiring new algorithms for each type of models and problem could arise if the user is unable to determine appropriate algorithms due to the lack of algorithms for that model or the difficulty in determining the type of models. Therefore, we propose an alternative of performing customization with the algorithm that is based on the task since the minimum requirement for a model to be useful is to know the task it is for. From among all possible tasks, we present customization frameworks and algorithms for classification and dimensionality reduction and perform experiments on them using real-world dataset. This experimental results show the effectiveness of our proposed methods.