A visualization of gendered collaboration among researchers funded by the DFG in order to explore gender equality within academia. To try the visualization, look no further than right here.
- Project description
- Design Decicions based on “Four Nested Levels of Visualization Design” by Munzner
- Validation
- Installation
- Manual
- Contributors
- Copyright
As one of the largest German instutition to fund research is commited to gender equality, it makes sense to look at their own data in order to analyze how and where measures need to be taken to own up to these promises. We're using parts of the GEPRIS data to look at collaborations within research project that are funded by the DFG. As research shows, people tend to form their networks with people who are similar to them, which can hinder gender inclusion in fields that have traditionally been shaped by male networks. To find out whether this is the case for German academia and research as well, we are visualizing the collaboration amongst researchers funded by DFG.
The target audience for this project are equal opportunity commissioners at universities and people interested in gender equality.
We're giving you the opportunity to explore different universities' degrees of gender equality to see where there's still work to be done.
Worldwide Gender-Name Dictionary
External libraries in addition to d3.js: jQuery, Waypoints.js
"There is now abundant data that diversity is essential to scientific excellence, with experiments showing that diverse teams have cognitive diversity that allows them to outperform homogeneous teams." reads a statement by Columbia University in 2018. While their statement focuses on students, the same applies to researchers.
We are thus interested in gender relations (as one aspect of diversity) within German academia. This is important knowledge for equal opportunity commissioners at German universities: How diverse are research projects regarding the gender ratios of their teams? How diverse are individual researchers' collaborations? And how diverse are universities regarding the gender ratio of projects that they host?
Our data is the GEPRIS dataset that contains projects funded by the DFG, including the people and institutions involved in those projects upon application. Please note that the GEPRIS data does not consider time, meaning that locations and titles represent the most current data. This could lead to bias and spurious links. We've also derived a probable gender for each person based on the first name. This data stems from Lax-Martinez, G., Raffo, J. and Saito, K. 2016. "Identifying the Gender of PCT inventors", WIPO Economic Research Working Paper 33.. We validated this assignment by picking the title "Professor" and "Professorin", and comparing them to the gender assignment. Less than 1% of the assignments were in contrast to the titles, meaning that we would assume a similar error ratio for the entire data set. Additionally, around 19% of first names could not be assigned a gender at all.
The original GEPRIS dataset is made up of tables of mainly nominal data (project IDs, person IDs, institution IDs, names, institution address...). We've created additional nominal data for an inferred gender and spatial data (geological position) for the institutions. Beyond that, we calculated a project_gender_index. This is value between 0 and 1, based on the gender ratio of all people involved in a specific project (that has more than one person). 0 indicates that there are only men in a project whereas 1 indicates that there are only women in a project. Following, the person_gender_index is the mean of project_gender_index for all projects a person is involved in and institution_gender_index is the mean of person_gender_index of all people related to that institution. Ordinal data consists of the amount of projects that a researcher is involved in (in order to find the top 10 researchers per institution) as well as that number summed up for each institution.
The tasks for this data are to discover distributions, browse the topology and identify & compare institutions.
We decided to explain a bit about the data we used and how we used it first in order to allow for critical thinking. Thus, some information is displayed in scrollable interaction before we display the main visualization. Said information is about the general gender ratio in the dataset (bar chart for quick orientation) as well as the ratio of correctly inferred gender (bar chart for the same reason). Additionally, there is a histogram to show the distribution of our project gender index. In bold text, at the end of each introductory paragraph for each visualization, we added the caption of the graph.
We decided to use a map to display the institutions in an overview and bar charts for detail. The map helps the user to find an orientation but also to get a first impression of the general distribution of data. The dot size for each institution represents the amount of projects that people belonging to that institution are involved in. The color of each dot represents the institution_gender_index. Users can zoom into the map in order to identify individual institutions. On click, we will provide further detail about the institution. Assuming that the researchers involved in the most projects can also have the most impact on future projects and their gender ratios, we show these 10 researchers per institution with their person_gender_index, using the same color scheme as before, and the amount of projects they are involved in. This provides an option to real life interaction in which researchers could be approached about strategies they are already using or that can help them diversify their collaborations - depending on said gender index.
Finally, we display a ranking of the top 30 institutions by their project count to illustrate how the most actively supported institutions are doing regarding their gender index. We're keeping the meaning of dot size and dot color from out previous visualization, thus providing linked highlighting.
More details on our design decision will now follow.
We added a transition animation to the first bar chart to capture the user's attention and to highlight the fact that this is a interactive visualization so that the user is invited to interact with the bar chart. On hover, we show the exact values for each category. With this bar chart, we also first introduce the user to the color scheme which we will stick to thereafter.
Color encoding: We tried various color schemes, but settled for blue - pink because it is widely used for color encoding gender and as such users will quickly and intuitively understand the meaning. Later on, we add a gradient to the color. We picked this specific blue and pink so the differences in the gradient are very pronounced and small changes are noticeable. We are, nonetheless, critical about how this color scheme reproduces gender stereotypes.
The stacked bar chart shows our validation of the data. The color scheme was chosen so the user quickly notices that "there is a lot of green", meaning most of the gender inference was correct. Green here representing that standard color for "good", yellow for "not so good" and red for "bad", which is very intuitive. The missing data was encoded in a neutral blue. On hover, we show the exact value in addition to the percentage.
The histogram links back to our color coding. Since the bins are double encoded with both the color as well as the gender index value, a legend for the color is not needed. The histogram illustrates how skewed the data is towards male domination of teams, while also giving an overview about the total distribution of the data. The text explains our "gender index" metric, as well as add some more basic facts about the distribution. We chose to keep it really light on the statistics, as not to overwhelm the user, and keep them focused on the main question.
The map invites the user to explore the data. This spatial representation facilitates searching and browsing: the user might know where the institution of interest is located, so they can quickly find them, much faster than e.g. a long alphabetical list. If they are interested in a specific institution, they may want to compare it to other institutions in that region: so the "interest" is space correlated. Lastly, a map is visually appealing and a welcome variety after displaying bar charts only, so far.
To prevent overlapping of the circles, we do not scale the circles when zooming in. The farther the user zooms in, the less overlapping occurs.
An overview-detail approach was chosen because while we do want to show the data about individual researchers, the researcher dataset is very large. We thus aggregated our data per institution and show this aggregated data on the map. Once the user clicks on an institution, they find out more detail specific to the top 10 most important researchers at the institution (measured by number of projects). In the "Top 10 researcher bar chart", the user can hover over the bars to see even more details about the person of interest. The "gender index" is displayed prominently for both the institution and the researchers in a consistent way: as a "badge" double encoded with the value of the index as well as the color.
We added a transition animation to show the linkage between this visualization and the previous map. The 30 most important institutions (again measured be the number of projects) from the map are rearranged in a sequential way, ordered by importance. This visualization was added to complement the map, to focus on the largest institutions and provide an alternative way of exploring the data. Here again, the user can interact and click on the institution to achieve the same results as with the map.
The dataset was already small due to the data preprocessing that we needed to create the gender index. However, we also decided to display only institutions that have more than five projects related to them. Thus, we could ensure a quick visualization with focus on those institutions that matter on a larger scale of this context.
We validated the design with one person who is doing their PhD in sociology and gender studies, and is thus very interested in this visualization. According to Munzner, possible threats are: addressing the wrong problem, bad data / operation abstraction, ineffective encoding / interaction technique, slow algorithm. We focused on the first three threats.
Because it help to clearly state the aim and give context as well as criticism about the data, the user appreciated the scrollytelling part of the visualization. They did find our questions relevant / addressing an important problem. However, they would have preferred a more critical discussion. First, gender is only one aspect of diversity, and the user did not see this fact reiterated to their satisfaction. Second, using a gender binary for analysis reproduces the erasure of the actual breadth of gender categories.
The user had further questions about both GEPRIS as well as additional data: Does the GEPRIS data allow for titles beyond the gender binary (e.g. "Professor*in")? Is the mismatch between title and inferred gender due to the German "generische Maskulinum"? How well does the external data handle non-German names? The user picked one non-German researcher and looked them up online, realizing that the gender inference for their name was faulty. However, we then understood that the user confused the color coding for a gender inference instead of the coding for the gender index, which is a different challenge to fix.
The questions could be addressed in the accompanying text. We could also add a link for each professor to the dataset in order to make it easier for the user to check the identity of each researcher and the correctness of the data. Additionally, we need to improve the color coding in the bar chart to make it more clear that the color coding is related to the gender index, not the researcher's inferred gender.
The user mentioned that the color encoding is very "classical" and thus intuitive, however, they wondered if a different color encoding could have been used in order to disrupt gender stereotyping.
The user very much enjoyed the last transition and scrolled back and forth between the map and the final ranking several times.
Luckily, our test user tried the visualization on a small screen and in a non-fullscreen window so that some of the visualization (e.g. sticky header) were displayed sub-optimally. Adjusting for different screen sizes and mobile would be an improvement.
- Clone this Repo
- run prototype.html via http
A hosted version is available here
Johannes P. and Laura L.
Data derived from original data provided by GEPRIS (c) Deutsche Forschungsgemeinschaft