Active Pattern Classification for Automatic Visual Exploration of Multi-Dimensional Data

Abstract

The practice of applying a classifier (called a pattern classifier and abbreviated as PC below) in a visual analysis system to identify patterns from interactively generated visualizations is gradually emerging. Demonstrated cases in existing works focus on ideal scenarios where the analyst can determine all the pattern types in advance without adjusting the classifier settings during the exploration process. However, in most real-world scenarios, analysts know nothing about data patterns before exploring the dataset and inevitably find novel patterns during the exploration. This difference makes the traditional classifier training and application mode less suitable. Analysts have to artificially determine whether each generated visualization contains new data patterns to adjust the classifier setting, thus affecting the automation of the data exploration. This paper proposes a novel PC-based data exploration approach. The core of the approach is an active-learning indicator for automatically identifying visualizations involving new pattern classes. Analysts thus can apply PCs to explore data while dynamically adjusting the PCs using these visualizations. We further propose a PC-based visualization framework that takes full advantage of the PC in terms of efficiency by allowing analysts to explore an exploring space, rather than a single visualization at a time. The results of the quantitative experiment and the performance of participants in the user study demonstrate the effectiveness and usability of the method.