12-02, 10:55–11:25 (Europe/Amsterdam), Auditorium
In this talk I hope to convince you that models are not either predictive or causal, but both perspectives should be combined to solve real world problems. I will use a concrete example of how we automate irrigation in greenhouses at Source.
The causal revolution has taught us that there is a world beyond generating predictions. We now known that not all ML models are suitable for causal inference and have alternatives like double machine learning for causal inference.
During this talk, I hope to convince you that the difference between predictive and causal models is not as clear cut as you might think. Using a concrete example of how we control irrigation in greenhouses using machine learning, I will give an example of how to break down a problem into model components that are more or less predictive or causal. Moreover, I hope to give you some practical guidelines on how to decide whether a predictive or causal approach is more suitable for the components of your model.
Outline:
- Causality warm-up
- Explanation of irrigation in greenhouses
- Demonstration of the caveat with predictive models
- Demonstration of why feature selection matters more then framework selection
- Brief introduction in double machine learning (full explanation is beyond the scope of this presentation)
- Demonstration of why double machine learning does not solve the feature selection problem
- Optional: link to Judea Pearl's causal graphs
- Explanation of how to isolate part of the problem where you can use predictive models
- Explanation how this components come together in our solution for irrigation control at Source
- General advice on how to identify whether to use a predictive or causal approach
- Conclusion
Previous knowledge expected
