On the Use of Parallel Coordinates for Temporal Multivariate Data
Kahin Akram Hassan*Link ¨oping University
Jimmy Johansson† Link ¨oping University
Camilla Forsell‡ Link ¨oping University
Matthew Cooper§ Link ¨oping University
Niklas R ¨onnberg¶ Link ¨oping University
Figure 1: Screenshots of 2D (left) and 3D (right) parallel coordinates displays, showing temporal data evaluated in this work. See Fig. 2 for detailed information regarding their respective interactive functionality.
ABSTRACT
This work presents the results from a user centered evaluation of visual representations of temporal multivariate data using 2D and 3D parallel coordinates. The objective of the evaluation was to inves-tigate whether 2D or 3D representations increase user performance when the data consists of temporal multivariate data and the visual representation contains interactive user tools. The results show that the 3D parallel coordinates representation outperforms 2D paral-lel coordinates with regards to both accuracy and response time. This result is of interest to the information visualization community, since it shows the usefulness of visual representations of temporal multivariate data.
Index Terms: 2D & 3D Parallel coordinates, temporal data, infor-mation visualization, evaluation.
1 INTRODUCTION
The parallel coordinates technique was invented by Ocagne [5], made famous by Inselberg [2] and introduced for analysis of mul-tivariate data by Wegman [7]. To represent a mulmul-tivariate item in parallel coordinates, a series of points are positioned on the parallel axes and are then connected by a polyline. Many researchers have made different approaches to improve upon parallel coordinates in order to reduce the cluttering effect when visualizing larger data sets, see [1] for a comprehensive overview. To achieve this, parallel coordinates has, in several works, been extended to three dimensions with the assumption that more variables can be displayed simul-taneously. For arbitrary multivariate data, previous research has concluded that standard 2D parallel coordinates are superior to 3D parallel coordinates for to the analysis of multivariate patterns [4].
The visualization of temporal data in parallel coordinates [8] has not received as much attention as visualization of multivariate data. However, several of the methods for reducing clutter are based on various 3D representations. When it comes to evaluation of parallel coordinates, surprisingly little work is found in the literature [3].
This poster adds to this body of work by presenting an evaluation
*e-mail: kahin.akram.hassan@liu.se †e-mail: jimmy.johansson@liu.se ‡e-mail: camilla.forsell@liu.se §e-mail: matthew.cooper@liu.se ¶e-mail: niklas.ronnberg@liu.se V1 V1 V2 T V2 T T
Figure 2: Illustrations of 2D and 3D parallel coordinates for temporal
data. V1and V2make up the relationship under study and T
repre-sents the temporal dimension. For 2D parallel coordinates (left), the temporal dimension can be investigated by using an interactive brush (illustrated by the semi-transparent gray box). For 3D parallel coordi-nates (right), the temporal dimension can be analyzed by rotating the 3D representation. The rotation was limited horizontally and vertically, to avoid unintended mirroring of the representation.
comparing 2D and 3D parallel coordinates for analysis of relation-ships in temporal multivariate data. The 3D version used in this work was originally proposed by Wegenkittl et al. [6], as illustrated in Fig. 2 (right).
2 TESTAPPARATUS
For the study 12 experts in visualization were recruited, with nor-mal or corrected-to-nornor-mal vision and no color vision deficits (self reported). These were highly experienced in the research fields of scientific and information visualization.
In the test, the participants had to identify one relationship in the visual representation on the computer display out of eight possibili-ties given on a reference sheet (see Fig. 3) on the desk in front of the display. These relationships were composed of two signals that vary over time. In 2D parallel coordinates the participant could use an interactive brush to the right of the visual representation (see Fig. 2 left), to scroll through the data to see changes in time, with an adjustable time. The dotted line shown in Fig. 2 was not present during the evaluation (as shown in Fig. 1 left), and is only shown here for illustration of the interaction. The reason not to connect the axes with the temporal dimension, was because this could be misleading for the participant since only the relationship between adjacent axes is normally shown in parallel coordinates. There exist many ways to design such an interaction, but the used interaction
Figure 3: The eight reference combinations on the reference sheet. Each combination was composed of two signals common in various domains.
technique was considered as natural according to participants in a pilot study. In 3D parallel coordinates the participant could rotate the visual representation to explore the data. Data was presented on the V1and V2axes and the T axis was used for the temporal variable
(see Fig. 2 right). The interactions were consequently different for the two representations. This was deliberately chosen since the aim of the study was to compare the best possible implementation of each technique. In addition, different colors were not used since the aim was also to evaluate the most basic versions of each technique for the data at hand.
Each test was initiated with a practice run for familiarization and to reduce learning effects. After this the test continued with 24 trials in each condition (2D and 3D parallel coordinates).
3 RESULTS
The results showed that the mean accuracy (see Fig. 4) for 2D parallel coordinates was 18.2 (SD=5.8), and for 3D parallel coor-dinates it was 22.6 (SD=2.4). When accuracy was analyzed with a repeated measures ANOVA with one within subject factor, visual representation (2D parallel coordinates, 3D parallel coordinates), a significant effect of visual representation was found (F(1,11)=10.87, p=0.007), revealing that the accuracy was significantly higher in 3D parallel coordinates compared with 2D parallel coordinates. The mean response time (see Fig. 4) was 43.7 seconds (SD=16.9) for 2D parallel coordinates, and 21.9 seconds (SD=9.1) for 3D parallel coordinates. When response time was analyzed with a repeated measures ANOVA a significant difference was found (F(1,11)=26.0, p<0.001), showing that the response time in 3D parallel coordinates was significantly faster than in 2D parallel coordinates.
4 DISCUSSION
For the study cohort, 3D parallel coordinates outperformed 2D par-allel coordinates in both accuracy and response time (see Fig. 4). The results show that for data exploration involving relationships formed by two variables that vary over time, 3D parallel coordinates is a more useful tool than 2D parallel coordinates.
For both visual representations, interaction was added to facilitate interpretation of the representations, and to simplify the task of recognizing the relationship in the data. The results suggest that interaction in 3D parallel coordinates was experienced as more intuitive compared to 2D parallel coordinates. The rotation of the representation in 3D always displayed all data, while the interaction in the 2D representation showed a selection of the data at different intervals in time. Consequently, the results might indicate that the loss of overview might have negatively affected the performance in the 2D representation. Furthermore, in the 3D representation the rotation enabled a visual representation close to a line graph, which was more closely linked to the combinations given on the reference sheet in contrast to 2D parallel coordinates, and this therefore further facilitated identification of relationships. Also, the interaction in 2D parallel coordinates involved a number of different mouse clicks, as compared to 3D parallel coordinates, which in turn might have taken longer time to perform.
5 CONCLUSION
The results show that for data exploration when having two variables that vary over time, 3D parallel coordinates is a better visual
repre-M ean p er formanc e (max=24) Accuracy 24 Error Bars: 95% CI p=0.007 20 15 10 5
2D parallel coordinates 3D parallel coordinates
Response time
2D parallel coordinates 3D parallel coordinates
0 0 10 20 30 40 50 60 M ean r esp
onse time (sec
onds)
Error Bars: 95% CI p<0.001
Figure 4: Left, the mean accuracy in recognizing the relationship in the visual representation (max=24) for 2D parallel coordinates and 3D parallel coordinates. Right, the mean response time in seconds for the two representations.
sentation than 2D parallel coordinates in terms of higher accuracy and faster response time. This is of interest to the information visu-alization community, regarding usefulness of visual representations of temporal multivariate data.
6 FUTUREWORK
For future work, it will be investigated whether the effect that 3D parallel coordinates outperforms 2D parallel coordinates remains for individuals with less knowledge and experience in visualization. With regards to the complexity with two variables that vary over time, it would be interesting to explore visual representations of one variable over time as well.
ACKNOWLEDGMENTS
This work was partly supported by the Swedish Research Council, grant number 2013-4939.
REFERENCES
[1] J. Heinrich and D. Weiskopf. State of the Art of Parallel Coordinates. In Eurographics 2013 - State of the Art Reports, 2013.
[2] A. Inselberg. The plane with parallel coordinates. The Visual Computer, 1(4):69–91, 1985.
[3] J. Johansson and C. Forsell. Evaluation of parallel coordinates:
Overview, categorization, and guidelines for future research. Trans-actions on Visualization and Computer Graphics, 22(1):579–588, 2016. [4] J. Johansson, C. Forsell, and M. Cooper. On the usability of 3D display in parallel coordinates: Evaluating the efficiency of identifying 2D relationships. Information Visualization Journal, 13(1), 2014. [5] M. Ocagne. Coordonn´ees parall`eles & axiales: m´ethode de
transforma-tion g´eom´etrique et proc´ed´e nouveau de calcul graphique d´eduits de la consid´eration des coordonn´ees parall`eles. Gauthier-Villars, 1885. [6] R. Wegenkittl, H. L¨offelmann, and E. Gr¨oller. Visualizing the behavior
of higher dimensional dynamical systems. In Proc. IEEE Visualization 1997, pages 119–125, 533, 1997.
[7] E. J. Wegman. Hyperdimensional data analysis using parallel coordi-nates. American Statistical Association, 85(411):664–675, 1990. [8] Y. Zhonghua and W. Lingda. 3d-parallel coordinates: Visualization for
time varying multidimensional data. In Proc. IEEE Software Engineer-ing and Service Science 2016, pages 655–658. IEEE, 2016.
