Evaluating the utilization of clustering methods connected with multivariate visualizations

Visualization software in all fields becomes increasingly interactive with more and more elements to support users. Elements that help an observer to interpret given data sets by computing certain properties of the data to enrich their visualization we call "visual scouts".

Many visualization programs implement visual scouts, but few is known about how users apply them in a real setting. In order to learn more about how users work with interactive visualization software we provide them with clustering methods (k-mean and hierarchical clustering), as one possibility to manipulate the data. Subsequently we evaluate how natural science students use these clustering methods as provided by an interactive comparative visualization software for high dimensional data (VisuLab®). The students are introduced to four different visualization methods (figure 1) and given the possibility to apply four clustering methods to help them when interpreting two sample datasets (table 1 first two rows). This is part of a guided instruction after which they independently interpret two new but related real data sets (table 1 last two rows). In total they work for about six hours.

During our evaluation we were able to record the clustering activities of ~250 students, by directly logging the user's activities within the software. Based on this data we computed how many times the data was clustered and with which visualization methods the results were visualized. In total we registered over 57'000 clustering activities of which 30% where recorded during the instruction period and the remaining 70% during the time they worked independently.

Based on the data gained we investigate relationships between clustering methods, the type of data analyzed and the visualization method chosen. This is shown in figure 2, where the number of clustering in regard to the visualization method is presented in a boxplot.