The dependency network viewer executes the stored procedure System.Microsoft.AnalysisServices.System.DataMining.DecisionTreesDepNet.DTGetNodeGraph
which yields some integer measure that represents the strength of influence of input variables on the output variable. How this measures are calculated (information gain, chi-square, correlation etc)? It seems as if conditional influences are not taken
into account: if A and B are two major factors which impact variable C, but A and B are strongly correlated so that C given A is not dependent on B, dependency algorithm will still depict B as the second major factor. Am I right? So there is no analogy
of tests like conditional chi-square, conditional mutual information or partial correlation?
It seems to me that MS Data Mining lacks some kind of Bayesian Networks algorithm wich would illustrate conditional dependencies. That would give
a useful insight on how various factors are related to each other and through what kind of chains a change in some input variable transforms to the output.
View Complete Post