Training Set Construction Methods

In order to build a classification or regression model, learning algorithms use datasets to set up its parameters and estimate model performance. Training set construction is a part of data preparation. This important phase is often underestimated in data mining process. However, choose the appropriate preprocessing algorithms is often as important as choose the suitable learning algorithm. Goal of training set construction algorithms is to build representative datasets by discarding useless instances and enforcing important instances. Good quality training set is a good premise to build a well learned and reliable model. Lot of literature have been published about comparison of learning algorithms and regression or classification models, but good review and comparison of training set construction methods have not yet been given. This work is focused on how to select data samples from an original set and place them into the training and testing sets. In the first part is an overview of existing approaches and new possible approaches are discussed. The second part is focused on experimental comparison of these methods.