The consideration of fairness in the development of machine learning-based solutions is getting traction as a key aspect of artificial intelligence and modelling of social behaviours. This week the Harvard Business Review published a story authored by the leading forces behind a health analytics project that is using Deep Learning to detect people with propension to have a cardiovascular condition.
The researchers behind the Stroke Belt Project have found some useful tips to those who want to avoid socio demographic, racial, or any other type of human bias introduced by those who either prepare the data, build the model, or optimized outputs.
In an effort to promote “Fairness by Design”, a concept that we, at BBVA Data & Analytics, try to have always present when modelling and sampling data, they prescribe a few advices:
- Pair data scientist with social scientist to introduce a humanist perspective to the data and the solution to be implemented.
- Annotate data with caution. Make those responsible for annotating data that is gonna feed a machine learning model aware of their own possible bias.
- Measure fairness. By utilizing metrics to measure fairness and you can correct bias.
- Don’t only focused on representativeness. Once fairness measures are in place find balance between representativeness of future cases and underrepresentation of minorities.
- Keep de-biasing in mind and if necessary, avoid sociodemographic categorization altogether.
Fairness is one of the main concepts that BBVA Data & Analytics has been working to introduce in financial modelling. We have published a recent paper, “Reinforcement Learning for Fair Dynamic Pricing (Maestre, 2018)”, about the importance of integrating fairness as principal design principle when relying on Artificial Intelligence to make financial projections (See Section B on the paper):
Moreover, we have proposed metrics to quantify fairness, as the HBR article points out (point 3 in HBR article):
We have also authored several articles, attended conferences to talk about this topic and help introduce this debate in the digital transformation that BBVA in undergoing. A special attention on groups segmentation must be paid (points 4,5 in the HBR article). We try to achieve this with our client2vec models.
As underlined in the aforementioned HBR article, introducing “Fairness by Design” is not a drag for innovation, but a way of making advanced analytics more solid, more reliable and with better result across different socio demographics segments.