6  Machine learning

6.1 Causality

https://youtu.be/-7vSiWRasxY

• ML learns outcome by finding distinguishing features and how cases differ
• There will be missing data… censored data.. bias…
• ML learns target outcome only as a function of the observable data
• People with expert/domain knowledge would be able to identify some of the people or fields missing
• ML also usually fails to account for relationships between variables
• In order to improve AI learning, need to incorporate causal knowledge
• Pearl’s ladder of causation: moving from association (what if I see) to intervention (what if I do) to counterfactuals (what if I had done)
• Argues that you can smartly use smaller datasets in causal models to get smart outcomes rather than just throwing big datasets in models and getting useless outcomes