Imprecision guidance
- Start w/ experts; ask what you would ideally like; think though intermediate measures; overlap science w/ the process
- Befriend the person who owns the data
Differing approach: Predictive v. Causal affects how you approach/change your systems
- Predictive: anticipating outcomes. Data that have to inform prediction of something that hasn’t happened
- Most AI today is predictive.
- Predicted outcome remains theoretical; it’s different from the actual. You decision doesn’t affect the probability
- Causal: Cause and effect prediction; my action is producing an effect
- Take action and try to foresee an outcome
- My actions influence the outcomes
Wind turbine example
- When you switch from 12 months to 6 months, the problem is you don’t have data for the switch. So you’re making an educated guess. To test, you will have to run an experiment.
- Digital twins?
- If… then… (these models weren’t built to answer a causal question) So the fan question is causal b/c we’re predicting that our changing the fan will prevent failing.
What’s the data I need to build a model?
- Input data
- Desired outcome
- Need history (dataset 1) > and potential/future (dataset 2)
Ford
- Linear regression, rule and looks at the Answer = y-axis at 0 + coefficient x the variable
- The ford model is a linear model. So it doesn’t matter where you are on the line because it’s just about the 700 (plug back into GPT)
Linear Probability Model
- “Feature” = x variable. “Target” = outcome (y variable)
But b/c the linear probability has negative #s and 100%+, we use an “ideal” probability prediction probability
- Impulse functions, b as large; b drops then it starts to look like a linear regression
- The greater the overlap, the larger of unexplainable surface area.
Business intelligence
- Profitable expectation as a way to create a screen/bar above which you must pass to submit/send
AUC is Rsquared’s cousin.