It’s interesting to study, but it doesn’t magically make hard problems tractable. It seems to make certain common mistakes less likely. When you do statistical analysis or machine learning to model p(x,y) or its friend p(y|x) it’s possible that you mistake correlations for causation (because there is no way to express directionality of dependency). If I have data about the measurement on a barometer and atmospheric pressure then I may learn a model that predicts higher pressure when the barometer reading is high. But if I manually force the barometer reading higher that doesn’t increase the atmospheric pressure. Causality provides a mathematical framework to express that idea.
The problem I have with causality is how to find accurate causal diagrams from unstructured observational data. We _could_ guess and check every possible causal relationship, but that’s at least exponentially hard—in which case causality is useless.
People seem to have reasonably good capabilities for generating candidate causal hypotheses from observational data (basically all of modern science), but most of the material I’ve found on causality focuses on the theoretical benefits it provides rather than on practical applications at scale. (I don’t care if we can automate finding the causal graph for barometer & air pressure; how do I find a causal graph for classifying fake/real news from plain text data?)
The problem I have with causality is how to find accurate causal diagrams from unstructured observational data. We _could_ guess and check every possible causal relationship, but that’s at least exponentially hard—in which case causality is useless.
People seem to have reasonably good capabilities for generating candidate causal hypotheses from observational data (basically all of modern science), but most of the material I’ve found on causality focuses on the theoretical benefits it provides rather than on practical applications at scale. (I don’t care if we can automate finding the causal graph for barometer & air pressure; how do I find a causal graph for classifying fake/real news from plain text data?)