bayeslite
bayeslite copied to clipboard
probcomp
Conditional dependence probability?
In addition to pairwise dependence probability, it would be great to predict conditional dependence probability; that is, the probability that two columns are dependent, given a third.
@axch suggests finding a mutual information implementation that will estimate from samples, for each column, computing mutual information with that and other pairs of columns, so that's an n^3 algorithm for full ternary dependence? I might easily have misunderstood. Need to read more.
This also apparently depends on the metamodel interface that @fsaad is working on, though I don't really understand what the dependence is. Perhaps being able to ask crosscat the question, rather than asking a separate mutual information engine the question?
Clarifications appreciated! This needs a design doc.
Please see issue #69 plus discussion which discusses conditional density. Once we have that, we should be able to extend to dependence. (ps sorry for closing, Close and Comment and Comment are close to one another...)
Related issue in CrossCat https://github.com/mit-probabilistic-computing-project/crosscat/issues/47
Given both dependent bugs are now fixed, @fsaad is it time to revisit this question? It would be pretty useful to start to tease out possible causal structures.
Those don't actually get us any closer to evaluating conditional dependence probability, I think...
