Lately, I’ve bumped into this article:
It describes how to discretisize continous variables throughout intervals, and give parents nodes weights, then compute probabilities for child node to be some value/interval. It’s something to make CPT (conditional probability table) less complex. I was wondering has anyone implemented suggested algorithm and also used pymc3?
Generally, i’d say its something like logistic regression, but weights are not computed, rather given from an expert in the field you’re constructing Bayesian network for.
Some similar ways to simplify CPT are Noisy-or, Logistic regression, CPTree etc.
1 Like