TL;DR: We conducted a user study to compare 3 geo-temporal visualizations for the analysis of propagation phenomena: small multiples, animation and a map with glyphs. Our results indicate that small-multiples performs better overall for this kind of analysis, but animation is better for identifying the direction of a propagation pattern and glyph maps are better to identify the arrival time of a disease in a particular location. Regardless of performance, participants felt more confident using animated maps.
Propagation phenomena occur in diverse domains, from epidemiology to cybersecurity. Examples include disease spread, keyword tags on social media, packet replication in communication networks, and software viruses, among others. In many scenarios, geography is key in understanding and analyzing the phenomenon. For example, in the health crisis we are facing with the COVID-19 virus, analyzing the impact of international trips and interactions using public transportation inside cities has allowed policy makers to apply restrictions and quarantines. In some cities, like Wuhan, these measures have helped to delay the dissemination of the outbreak and reduce the number of infected people.
But, how can we design an effective visualization to support his task? To answer this question, we evaluated three geo-temporal visualizations in a user study in which participants had to describe geo-temporal patterns. To design our study, we considered the following 3 questions:
- What are important questions for analyzing geo-temporal propagation patterns?
- Which visualizations are best for helping people see these patterns?
- What are realistic scenarios for evaluating propagation visualizations?
What are important questions for analyzing geo-temporal propagation patterns?
Our first task was to identify what are the questions that researchers ask when analyzing such data. We reviewed related literature and found that, for example, historical epidemics often spread in a wave-like fashion between contiguous regions. Such is the case of the Black Death in the 14th century, that spread in Western Europe as a wave of infection at a more or less constant speed for three and a half years. Today, however, we are much more connected than when these epidemics happened, especially in terms of transport. In particular, planes now allow diseases to spread with “hops” between locations and not necessarily between contiguous regions (see Figure 1). We observed these hops in the way COVID-19 spread, as it initially started in China, but quickly reached the United States, Italy, France, Australia, and other countries distant from one another.
Our review led us to 5 dimensions to characterize a spatio-temporal propagation pattern:
- Hops: Indicate if a propagation phenomena takes place between adjacent locations, or if geographical jumps occur.
- Scope: Indicate if a disease propagates across an entire geographical region or not.
- Direction: Indicate if the virus spread in a particular direction or in multiple directions.
- Arrival: Indicate how much time it takes to the virus to propagate to a particular region.
- Peak: Indicate when the propagation reaches its maximum value in a single region.
Researchers usually analyze propagation patterns considering all these dimensions at the same time. For example, a disease can spread with hops in multiple directions. However, we decided to evaluate them individually so the results of each dimension wouldn’t interfere with each other. That is, we asked our participants to describe a pattern one dimension at a time and designed scenarios in which that dimension was as clear as possible.
Which visualizations are best for helping people see these patterns?
After identifying a set of characteristics to describe a geo-temporal pattern, our next step was to select which visualizations would be appropriate to see them. Available alternatives in the literature can be broadly divided into two main categories: those that juxtapose location and those that juxtapose time. In the first category we find visualizations like small multiples that represent each time-step (years, months, days, etc) as a map positioned one next to the other. In the second category we find visualizations like space-time cubes where the plane in the horizontal axis displays a map with the locations to study and the vertical axis displays time.
We considered a broad range of techniques, several of which we discarded based on inherent limitations. For instance, space-time cubes, in which temporal information is encoded in the vertical axis while the geographic is encoded in the other two, were not considered as they typically yield cluttered views. We ultimately chose visualizations from each category that were commonly used:
- Small multiples (juxtapose location): Each time-step is represented as a choropleth map, displaying all maps simultaneously by juxtaposing them in a grid pattern (Figure 2-a).
- Animation (juxtapose location): Each time-step is represented also as a choropleth map in full screen. Each time-step corresponds to one frame of an animation (Figure 2-b).
- Map with glyphs (juxtapose time): This technique superimposes glyphs on top of a base map. Each glyph is associated with a geographical region, and encodes time-steps as small cells organized in a grid pattern (Figure 2-c).
What are realistic scenarios for evaluating propagation visualizations?
The final step in our study design was to create scenarios that would mimic real world propagation phenomena. We had two options: consider real world data or generate our own datasets. The first option has the advantage that it would have allowed us to evaluate them directly in a real world context. However real world datasets didn’t offer us enough flexibility to create patterns to evaluate the characterization of patterns by the 5 dimensions individually. For this reason we finally decided to generate artificial datasets to control the different propagation parameters and ensure homogeneous task difficulty and the remaining question was, how to ensure reliability of these datasets?
Fortunately, simulations in epidemiology have been used for decades, giving us a solid framework to generate datasets realistic enough for our purposes. These simulations required two main ingredients: a way to model a population, and a way to model the disease itself and its spread on the population. For the population, we used a network in which each individual is represented as a node, and where two nodes are connected if they are at a minimum distance between each other. We used the NetworkX library to generate and manipulate these networks.
We modeled the disease with the SIR compartment model, which is commonly used, easy to setup, fast to compute and freely available. The model works as follows: each individual in a population can have three possible states: S, susceptible to a disease; I, infected, can spread the disease to others; or R, recovered or died, and therefore no longer part of the simulation. Once the simulation starts, infected persons (I) can pass the disease with a defined probability to susceptible individuals (S) connected to them. Infected individuals (I) can also recover (R) with a defined probability. The infection and recovery probabilities are set between 0.1 and 0.5, depending on how fast or slow we want the disease to propagate. We used the epydemic Python library for the simulation of the disease propagation.
Finally, we needed to select the geographic regions where our simulations would be executed and the length of the time interval they would represent. To prevent participants from trying to guess answers based on their familiarity with a particular region, we used three distinct maps from regions that they were unlikely to identify. The maps were chosen to have a range of diverse configurations in terms of division count and density (as shown in Figure 3). For example, we can observe in Figure 3 that the map of Biobío has a lot of regions with more regular sizes than the ones from Zacatecas. We tried to generate long simulations that each would represent a time interval of 35 years.
Which visualization performs best?
We conducted our study with 18 participants. Each dimension was used as a question with 6 repetitions each. We collected the time participants took to complete each task, the number of errors they made, and how confident they felt when answering them. Additionally, we asked them to rate the three visualizations over all the tasks.
Our results indicate that the effectiveness of each visualization depends on the task considered, but small-multiple maps perform best overall. More specifically, small-multiple maps are faster than animated maps and glyph maps but there is no difference among techniques in terms of error rate, at least when all tasks are considered collectively.
Interestingly, performance metrics are not aligned with subjective ones. Animation obtained the highest self-reported confidence score for almost all tasks, regardless of performance. This preference also shows in the overall ranking, as animation was preferred by 2/3 of all participants.