Learning-Based Dissimilarity for Clustering Categorical Data

Comparing data objects is at the heart of machine learning. For continuous data, object dissimilarity usually taken to be distance; however, for categorical there no universal agreement, categories can ordered in several different ways. Most existing category measures characterize distance among values an attribute may take using precisely number takes (the space) and frequency which they occur. These kinds overlook interdependence, provide valuable information when capturing per-attribute dissimilarity. In this paper, we introduce a novel measure that call Learning-Based Dissimilarity, comparing data. Our characterizes between two given terms how likely it such are confused or not all dataset with remaining attributes used predict them. To end, algorithm that, target attribute, first learns classification model order compute confusion matrix attribute. Then, our method transforms into measure. We have successfully tested against 55 datasets gathered from University California, Irvine (UCI) Machine Learning Repository. results show surpasses, various performance indicators clustering, most prominent relations put forward literature.