Calibrating for Gender Bias in Online Data

Two hispanic women laughing and smiling on a park bench

As companies and governments become more data-driven, there is one big problem: Are they designing the world for men?

The gender data gap refers to the overrepresentation of men in much of the data that organizations use to make decisions. While "Invisible Women" by Caroline Criado Perez demonstrates how this bias affects all walks of life, it is certainly present in online data, including the types of data that Citibeats' clients use.

Online platforms are often not representative of the gender distribution in the general population, which in most countries is 49% to 51%. For example, women represent 38% of Twitter users globally, with variations across countries. In some contexts, men also post more than women. While this varies from country to country and topic to topic, we often find around 70% of the Twitter conversation is male and 30% is female. 

Clearly, there is a discrepancy. We’ve made it one of our main goals to remove this bias and calibrate the results for our clients. Here is the why and how.


Why the Gender Data Gap is a Problem

At Citibeats, we see differences in the topics discussed by women and men. By calibrating for this effect, we give an equal weight to both. 

For example, in one sample of data on civic opinions in Latin America during the COVID-19 crisis, there is an underrepresentation of concerns about the healthcare system, household economy, and civic initiatives. That’s because women put more focus on these topics than men, but they weren't given enough importance in the data.


While we see slight differences after calibration (health system, household economy, and citizen initiatives have higher weighting), we don’t see major differences. It should be noted that this is averaged data from many countries and must also be analyzed by country. That said, it is important to carry out this type of analysis and calibration because in some cases, the underrepresentation could be greater and the calibration would have a higher impact.

We also observe country-specific gender differences. In the same data sample, we see that men in Brazil often discuss the business economy and health system issues, while women focus more on mental health and education.

In that case, mental health concerns would have a higher percentage of the discussion relative to other issues if women’s voices were weighted equally to men's.

Chart for proportion of conversation on civic topics during Covid-19

Visibility of gender differences also helps detect emerging issues. In another client application of Citibeats focused on consumer protection in three African countries, we found that COVID-19 exacerbated certain differences. In one country, the number of women reporting being victims of fraud increased more than it did for men. In another country, the number of women reporting being mistreated by customer service increased more than it did for men. 

By weighting results between genders and highlighting gender-specific problems, we can limit the bias of the gender data gap and give our clients the tools they need to make important decisions.


How Citibeats Bridges the Gap With AI

To calibrate results, we’ve been working on state-of-the-art technical approaches to infer gender from online discussion. We set out to understand, at the aggregate level, if posters on a topic are male or female.

In order to estimate the gender of a user, we focus on using people’s names, and for Twitter, bios. Using deep learning, our system looks for clues and makes a final determination of gender probability. A name like Esther may have a 100% probability of being female, while a name like Cris might have a 75% probability. From the bio, we may detect other clues, such as “mother of two,” “she/her,” or “empresaria” (this last example, "business woman" in Spanish, appears in Latin languages with gendered nouns). 

We found many interesting clues along the way. Women tend to use emojis more than men, and the use of male and female emojis can imply a user's gender. All these small discoveries are factored into the probability that our algorithm calculates. 


An Ethical Process for an Ethical Output

Ethics have been an important consideration in our approach. We want to limit the gender bias without creating privacy concerns. With this in mind, we’ve taken the following steps:

  • We only work with the person's name and bio to estimate gender. We don’t analyze photos, followers, or any other information, which we subjectively judge as more intrusive. Interestingly, we manage to reach the same level of precision and recall as state-of-the-art approaches that use these other variables.
  • We do not store user data. We only use it to train the initial model and delete it when it is no longer needed.
  • We only display the gender data as aggregated and anonymous. It is not possible to know the gender of a given individual, only the breakdown at a macro level (minimum 1,000 people). This means it ensures representative results without targeting individuals.

The Journey to Ethical AI

Our focus at Citibeats is grounding ethical AI in actionable, practical measures. We aim to balance idealism and pragmatism by putting concrete measures in place to apply ethical AI to social good challenges worldwide.

Ethical tools can help solve meaningful problems. If you have an idea for how to apply analysis of people’s opinions at scale with your organization, please contact us