The phrase Exploratory Data Analysis can have many meanings, in this context, it means assess the degree of association between a target and a large number of predictors.
ExploratoryDataAnalysis
calculates the following metrics from
the contigency table of predictor vs. target:
-
Mutual Information: Kullback-Liebler Divergence of the observed probabilities from the conditionally independent probabilities constructed from the observed row and column marginals
-
Phi coefficient: ϕ is sqrt(χ² / n)
If target is binary,
- Information Value: Symmetric Kullback-Liebler Divergence between the Class 1 distribution and Class 0 distribution
Computationally, entropy based metrics such as Mutual Information and Information Value need to take care of 0 probabilities as they result in Infinite entropy. Many literature and implementations add a small positive number to the frequency table to avoid log of 0, this is because the software isn't capable of dealing with infinities. Julia, however, does handle infinities gracefully, thus this package use Infinities when there are 0 probabilities. To avoid infinities and also not artificially adjust probabillities, re-bin the data so that there are no 0's in the frequency table.