Automatic KDD Data Preparation Using Multi-criteria Features

摘要

We present a new approach for automatic data preparation, applicable in most Knowledge Discovery and Data Mining systems, and using statistical features of the studied database. First, we detect outliers using an approach based on whether data distribution is normal or not. We outline further that, when trying to find the most appropriate discretization method, what is important is not the law followed by a column, but the shape of its density function. That is why we propose an automatic choice for finding the best discretization method based on a multi-criteria (Entropy, Variance, Stability) analysis. Experimental evaluations validate our approach: The very same discretization method is never always the most appropriate.