Force Feature Spaces for Visualization and Classification

Distance-preserving dimension reduction techniques can fail to separate elements of different classes when the neighborhood structure does not carry sufficient class information. We introduce a new visual technique, K-epsilon diagrams, to analyze dataset topological structure and to assess whether intra-class and inter-class neighborhoods can be distinguished. We propose a force feature space data transform that emphasizes similarities between same-class points and enhances class separability. We show that the force feature space transform combined with distance-preserving dimension reduction produces better visualizations than dimension reduction alone. When used for classification, force feature spaces improve performance of K-nearest neighbor classifiers. Furthermore, the quality of force feature space transformations can be assessed using K-epsilon diagrams.