DualRadviz: Preserving Context between Classification Evaluation and Data Exploration with RadViz

Abstract

With today's flood of data coming from many types of sources, Machine Learning becomes increasingly important. Though, many times the use of Machine Learning is not enough to make sense of all this data. This makes visualization a very useful tool for Machine Learning practitioners and data analysts alike. Interactive visualization techniques can be very helpful by giving insight on the meaning of the output from classification tasks. This also applies to the data itself, as visualization can make some characteristics of the data become clear. Several bi-dimensional projection methods have been used to visualize data instances based on their attribute values. This visualization is more difficult when the instances have a large number of attributes. One of the visualization techniques that can deal with high dimensional data is Radial Coordinates Visualization (RadViz). RadViz can also be employed to visualize the performance of a probabilistic classifier, helping a user to find problematic instances that might have been misclassified. In this study, a new approach to use RadViz is proposed and investigated. The proposed approach combines the two aforementioned uses of RadViz (attribute-based data exploration and result exploration based on the output of probabilistic classification). For such, it approach provides an easy transition between the two types of visualization. This allows the context to be preserved, since the user can visually track the same data instance from one type of visualization to the other. In order to evaluate the proposed approach, a prototype, named DualRadviz, was implemented. On this prototype, in addition to RadViz, visualization by Parallel Coordinates is also provided, so that precise instance inspection can be performed, since, different from RadViz, Parallel coordinates visualization does not suffer from ambiguity. To illustrate the usefulness of the proposed method, a case study is presented.