Using predictive analytics for social impact: ATB competes in the Women in Data Science Datathon.
Sometimes the pressure and excitement of a team competition brings out the best in people. We know that our data scientists at ATB are amongst the brightest in the world, so when we have a chance to compete against some of the top teams, we don’t think twice.
Most recently, four ATB team members competed in the Women in Datascience (WiDS) Datathon, predictive analytics challenge hosted by Kaggle and put on in a joint effort between the Institute for Computational & Mathematical Engineering (ICME) at Stanford University, Intuit, West Big Data Innovation Hub, and InterMedia.
The competition seeks to encourage teams (that must be 50 per cent women) to apply machine learning and statistical modeling in a predictive analytics challenge.
In this instance, teams were given a dataset with over 1200 features on demographics and behavior and were tasked with building a model to predict the gender of survey respondents. Read more about how the predictive models are intended to help design financial services for those excluded from the formal financial system:
There were 231 teams made up of 512 competitors from across the world. Teams were provided a training data set, which comprised 16 per cent of the entire data. In preliminary submissions running models on the training data, ATB was ranked second amongst the 231 competing teams.
Final rankings were tabulated when the predictive models were run on remainder of the dataset not used for training; ATB ultimately finished 17th.
Along with the thrill of a high pressure competition and the excitement of being ranked second leading up to the finish - there were important learnings that the team took away from the competition.
After almost 80 models tested by team members, the final ATB submission was a weighted average of the top three models developed by the team - a real group effort.
Beyond the team building and the excitement of getting ATB on the map within the the world’s largest machine learning community, the model and the techniques used to develop it could potentially be applied to other challenges.
Though predicting binary gender is not an exact solution that ATB would be implementing, the challenge was fundamentally a classification problem. A similar predictive model could be used for predicting fraud, or other attributes that might help ATB team members provide optimal service to customers.
Tackling the classification problem also gave ATB data scientists an excellent opportunity to test the machine learning technique of extreme gradient boosting on a real data set. This highlights an exciting outcome of participating in competitions: the opportunity to discover and use tools and techniques in the developing field of machine learning.
A big shout out to the ATB WiDS team, Samiul Azam, Yukun Zhang, Michael Akinwumi, Gunjan Kaur. We’re looking forward to future opportunities to showcase the talent here at ATB, learn and compete with other top data science teams from around the globe, and apply our learnings to bring the best products and services to ATB customers.