Building Prediction Models for AI Services: with Astrid Enslev Vestergård (DK) and Pia Thomsen (DK).
In this masterclass, you will get an insight into how data scientists have worked with predicting problematic gambling behaviour, based on findings from this past spring, when they modelled Danske Spil’s Game Scanner. The Game Scanner is an algorithm that identifies online gamblers with potential problematic behaviour. The modelling output is used by Danske Spil’s advisors on responsible gambling as an entry point to having telephone conversations with customers about a very sensitive subject, namely that of gambling habits. The sensitive use of the output restricted the model and feature choices. However, the expertise of the end user also created a possibility for feedback of the model performance that helped improve the model.
Participants will take away:
- Reflections on how to work with unbalanced data.
- Selecting the right model for the job.
- Seeking expert and user involvement.
Pia Thomsen (DK)
Pia holds an MA in Mathematics and Economics from the University of Aarhus. She has worked 7 years as an actuary at If Skadeforsikring before she joined Danske Spil in 2018.
Astrid Enslev Vestergård (DK)
Astrid holds an MA in Engineering, Mathematical Modeling and Computing from the Technical University of Denmark. Before joining Danske Spil, she worked for two years as a data scientist consultant at SAS institute. She also joined Danske Spil in 2018.
How did they work with AI?
Much of the work they have done with this model has been with figuring out what data to use and how to transform it. They have had the extra challenge that the core system at Danske Spil was replaced last year, changing the data structures so the raw training data no longer had the same structure as the raw production data. This meant they had to consider this when creating features for the model. They have tested different algorithms for the project and ended up with an xgboost model and used SHAP (Shapley Additive explanations) values for better explanation. Xgboost is an ensemble-tree based model and SHAP values are feature attribution method based on expectations. Until the model output is a UI layer for the users. They have used power bi to create an interface for the consultants who are performing the care calls. Care calls are the preventive calls they do when a person has been identified of the model. The consultant will have a talk with the person about the gaming habits identified and, if the person wishes, the possible options to cut back or stop gambling entirely.
Check out all the Techfestival Masterclasses.
17:00 – 18:00: Distributed Design, Values and Application.
17:30 – 18:30 Gig Economy
18:00 – 19:00 Public Code in Government
17:00 – 18:00 Driving Digital Transformation in Gastronomy.
18:00 – 19:00 Building Prediction Models for AI Services.