An Interactive Bio-inspired Approach to Clustering and Visualizing Datasets

Abstract

In this work, we present an interactive visual clustering approach for the exploration and analysis of datasets using the computational power of Graphics Processor Units (GPUs). The visualization is based on a collective behavioral model that enables cognitive amplification of information visualization. In this way, the workload of understanding the representation of information moves from the cognitive to the perceptual system. The results enable a more intuitive, interactive approach to the discovery of knowledge. The paper illustrates this behavioral model for clustering data, and applies it to the visualization of a number of real and synthetic datasets

Interactive Clustering

“…In Guo et al [49], users incorporate their knowledge of the data by identifying potential clusters and refining them through the attractive/repulsive operators and by joining and deleting clusters in a graph view. Similarly, Erra et al [43] and Awasthi et al [9] allow splitting and merging the clusters; in particlular, Erra et al [43] allows the users to interact with the data points that can yield immediate clustering results using GPU. In both Choo et al [31] and Sourina and Liu [93], the users are enabled to compare and specify various clustering parameters, enabling them to find the best clusters based on their domain knowledge.…”

“…We find that most of the approaches support updating the usual parameters of the clustering methods, e.g., adjust the number of clusters or the similarity threshold parameters [5,8,19,38,43,45,49,62,65,68,70,72,79,[92][93][94]101]. A somewhat unique perspective in this category is provided by Xiao and Dunham [111], where the authors note that when the whole dataset is unknown a priori, choosing an appropriate value of the number of clusters may be difficult.…”

“…The split and merge commands from a user can directly modify the structure of clusters [10,13,43,66,67] without any specification of how a cluster should be split. In Choo et al [30] and Hoque and Carenini [53], the users incorporate their domain knowledge through splitting and merging topics (modeled as clusters) according to their preferences.…”

“…Looney [70] uses a modified Xie-Beni validity measure to assess compactness and separation among clusters. Erra et al [43] performs a quality test on their method using precision and recall. They find that precision and recall values were higher than those compared with k-means and hierarchical approaches.…”

“…In Xu et al [113], a TableLens visualization can be used to zoom in on details of individual nodes in the overall clusters in bipartite graphs. Additional visualizations that have been used to support this functionality are scatter plots and dendrograms [31,88], histograms, bubble charts and chord diagrams [8], box plots [75], matrix-based visualizations [43,82], and parallel coordinates [31,98].…”