Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

Update 29/06/2017: The approach outlined in this blogpost was taken under the assumption that gender is binary but that is incorrect. Do read on as it might inspire you to do better and if you do wish to contribute to the topic check out Forwards.

The representation of women in technology is gaining more and more attention. In the R community, R-Ladies has grown massively over the past year and been awarded funding by the R-Consortium. The useR! conference has offered diversity scholarships (for underrepesented minorities in general) for the conferences in 2016 and 2017. This is a good thing but it does make me wonder: how do we measure success? how will we know these actions have lead to a better situation? And by how much?

This blogpost by Maelle Salmon and this one by Colin Fay inspired me to explore this issue as follows: if we could somehow measure the representation of women currently in our community we can keep track of this. As Mango are organisers of the EARL conference it would also help us assess our own performance.

Before I jump to the code I should mention that this is a simplified approach to the problem. The issue of equal representation is much more complex and the subject of a separate branch of research. This blogpost is an attempt to approach the problem of data collection and measuring progress from a different angle.

So I set out to measure the proportion of women among the faces of #rstats. I used Colin’s code to first retrieve the profile pictures.

# Taken from http://colinfay.me/playing-with-rstats-and-microsoft-computer-vision-api/
library(dplyr)
library(rtweet)
token <- create_token(app = "rstats_faces", consumer_key = CONSUMER_KEY,
consumer_secret = CONSUMER_SECRET)

users <- search_users(q= '#rstats', n = 1000, parse = TRUE) %>%
unique() %>%
mutate(profile_image_url = gsub("_normal", "", profile_image_url))

Then I used the Microsoft Face API to tell me if the picture was of a man or a woman. Colin used lapply but I thought I’d give purrr::map a try.

library(httr)
library(purrr)
getGender <- function(myURLs){
requestURL <- "https://westus.api.cognitive.microsoft.com/face/v1.0/detect"
parameters <- list(returnFaceAttributes="gender")

result <- map(myURLs,
function(myURL) {
Sys.sleep(3)
POST(requestURL, query=parameters,
'Ocp-Apim-Subscription-Key'=FACE_API_KEY),
body=sprintf('{"url":"%s"}', myURL)) %>%
content("parsed")
}
) %>%
# check if a face was detected and no error was returned
map_chr(~ ifelse(length(.x)>0 && !("error"%in%names(.x)), .x[[1]]$faceAttribute$gender, NA))

return(result)
}

usersWithGender <- users %>%
mutate(gender= getGender(profile_image_url)) %>%
select(user_id, gender)

From Maelle’s great collage I already knew not all profile pictures are of people and the Face API does not return any predictions in these cases. There are also some instances the Face API wrongfully didn’t detect a face (sorry Hillary Parker). Altogether the Face API detected 314 faces and the proportion of women among the #rstat faces is 20.4%.

This value is our first estimate and to assess its accuracy I want to compare it with other estimates. I have tried my best to find other sources but that has proven harder than I thought. Below I have plotted my estimate with estimates based on Dutch labour statistics1, Google diversity statistics2 and the StackOverflow Developer Survey3.

Although the percentage of women in technology at Google and in the Netherlands are based on an entirely different populations our estimate is close to these two. I have no specific knowledge about the #rstats community that would make it different from the wider tech community so I would expect the estimates to be close to each other. However there are some issues with the general estimates. The industry categorisation of the Netherlands means that we’re not capturing all tech jobs (for example, those in finance aren’t included). Also there has been quite the controversy on Google’s gender gap which may mean that estimate is too low as well. And finally in the discussion on the results of the StackOverflow Developer Survey it is stated their estimate is probably too low.

If I knew about these issues then why did I use them in the first place? Because it’s the best I could find and that is a problem. Without an accurate measure and without knowing how accurate my own measure is, it is clearly impossible to use it as an KPI 4. I will come back to this point later on but first I want to show how this methodology to measure the proportion of women can still be useful.

### Let’s go to the conference!

Our approach so far has been to use profile pictures, estimate the gender and then calculate the percentage of women. By applying this approach to conferences we can semi-automatically estimate the percentage of women for these conferences and then compare them with each other and year over year. The conferences we will be looking at are EARL and useR!.

First, we have to do some webscraping to retrieve the profile pictures. This is relatively easy with the rvest package. Below is the code for EARL London 2016.

library(rvest)
earl2016URL<- "https://earlconf.com/2016/london/speakers"
# We used the SelectorGadget to determine the name of the xml node containing the image tage
html_nodes(".col-md-3 img") %>%
html_attr("src")

earl2016Genders <- getGender(paste0(earl2016URL, speakers))

For the EARL conferences of 2017 the above code doesn’t work because the website was updated. Unfortunately for our sys admin I know where he lives and he helped me figure out how to get the image URL’s for all the speakers. It involved some wrangling with json but luckily we have the jsonlite and purrr packages for that.

library(jsonlite)
earlLondon2017Genders <- earlLondon2017 %>%
map_chr(~getGender(.))

earlSF2017Genders <- earlSF2017 %>%
map_chr(~getGender(.))

I wasn’t able to find profile pictures of the speakers for useR! 2017 but for useR! 2016 I was able to use almost the same code as above.

user2016URL <- "http://schedule.user2016.org/directory/speakers"
html_nodes("img") %>%
html_attr("src") %>%
map(~paste0("http:", .x)) %>%
map_chr(~getGender(.))

Finally, armed with the web scraping results we can plot the gender distribution for different conferences.

![](faces-of-rstats-002.png)

Compared to 2016 the percentage of women speakers at EARL has gone up. This reflects the effort that has been put in by the EARL team and will hopefully continue to improve in the future. The estimate for useR! 2016 is surprisingly close to our estimate for #rstat faces. The organisation for useR! 2017 will probably improve on this considering they’re providing childcare, diversity scholarships and the location (who doesn’t want to go to Brussels!).

### Summary

I’m just scraping the surface of gender prediction with this blogpost (see for example the genderizeR package for another approach). Also there are a number of caveats to my analysis that I didn’t cover such as the accuracy of the Face API, the bias in twitter users or the bias in conference attendees. This blogpost wasn’t intended to be at the level of an academic paper but rather a short exploration into a different approach to the problem.

As such my conclusion is that my method is definitely not complete. Without ground truth values it’s hard to say how accurate my #rstat faces estimate is. The values that I did collect weren’t consistent but that doesn’t mean there aren’t any better statistics out there (if you’re familiar with any please let us know!). Furthermore, as a standalone KPI this estimate won’t suffice but it could be useful in other contexts together with other values. The conference estimates are an example of this.

There are many more things that I wanted explore but unfortunately there was no time. If you’re interested, the code for this post can be found on GitHub and you can always reach us via twitter.

1. CBS.nl / Arbeidsvolume naar bedrijfstak en geslacht / 62-63 IT- en informatiedienstverlening

3. Developer Survey Results 2017 (https://insights.stackoverflow.com/survey/2017#demographics)

4. I don’t give up that easily. Another approach to validating my estimate is to count the members of all R-Ladies groups. After correcting for duplicates this might give me a more accurate estimate but it is beyond the scope of this blogpost.