Скачать или смотреть BayesiaLab Independence of Causal Influence Models

One of the most challenging tasks when building Bayesian networks with a group of subject matter experts is the parametrization of the model, i.e., the quantification of the probabilistic relationships. Indeed, with the default tabular representation of Conditional Probability Distributions (CPDs), the number of probabilities to be elicited grows exponentially with respect to the number of parent nodes.

As a result, there is not only a problem regarding the time it takes to fill in these large Conditional Probability Tables (CPTs) but also a problem of consistency between the elicited probabilities.

Even though formulas offer an effective way to fill CPTs, it is not really possible to use them "live" in a brainstorming session because they can seem quite complex and abstract to experts. An elegant alternative approach is to determine whether the model requires consideration of all interactions between the parent nodes. If not, we can "divorce" the parent nodes and effectively represent their Independent Causal Influence (ICI) on the child node.

BayesiaLab 10 includes a new tool for automatically generating different types of ICI models. Experts can choose the Combination Function (Or, And, Max, Min, Sum, Vote, Average), and then elicit Local Causal Effects using simple counterfactual questions.

ICI models are not only useful for modeling expert knowledge but can also be useful in the context of Supervised Learning. The model described above is indeed a Bayesian Regression Model. When data is available, Bayesian Parameter Updating can be utilized to automatically estimate the Local Effects.