Building Conviction and Credibility in ML-based Solutions – Armin Catovic, EQT

Session Outline

It is not enough to build best-in-class data pipelines and machine learning models. The insights from the ML based solutions are ultimately being consumed by your users. In this presentation we go through our journey in trying to instill confidence and credibility in our ML driven insights, within the context of private capital. We discuss the role of active learning, scoring models and explainability, data fusion, and visualization methods, in order to better engage our users (investment professionals).

Key Takeaways:

  • Active learning methods can be used to better incorporate the user within the ML pipeline, and improve their trust, i.e. human-in-the-loop
  • Extra care is needed when building complex scoring models – consider ranking/guiding vs decision making; the ability to quickly provide explainability and feature importance is very important
  • Understand your users and their business context, and adapt how you display and visualize insights accordingly
  • Continuously iterate over data quality – you can never get it perfect, but there are techniques to build more trust in data quality by e.g. fusing different data sources

Add comment