Visual approach to boundary detection of clusters projected in 2D space

Abstract

Data mining tasks are commonly employed to aid users in both dataset organization and classification. Clustering techniques are important tools among all data mining techniques because no class information is previously necessary – unlabeled datasets can be clustered only based on their attributes or distance matrices. In the last years, visualization techniques have been employed to show graphical representations from datasets. One class of techniques known as multidimensional projection can be employed to project datasets from a high dimensional space to a lower dimensional space (e.g., 2D space). As clustering techniques, multidimensional projection techniques present the datasets relationships based on distance, by grouping or separating cluster of instances in projected space. Usually, it is difficult to detect the boundary among distinct clusters presented in 2D space, once they are projected near or overlapped. Therefore, this work proposes a new visual approach for boundary detection of clusters projected in 2D space. For that, the attributes behavior are mapped to graphical representations based on lines or colors. Thus, images are computed for each instance and the graphical representation is used to discriminate the boundary of distinct clusters. In the experiments, the color mapping presented the best results because it is supported by the user’s pre-attentive perception for boundary detection at a glance.