The nutcracker framework for ensemble interpretability

The basic principles behind ensembles (e.g. Random Forest, AdaBoost) are simple. But we’re still in trouble when attempting to explain the logic taken. Where does the problem lie? The reason that ensembles are effective is that the base estimators work together and compensate each for the others’ shortcomings. The Nutcracker Framework Given a trained ensemble and the relevant training / test dataset, construct prediction matrix, M, cases (rows) against predictions (columns). Bicluster M to a given number of R x C biclusters. Now, investigate performance per bicluster (R x C). Identify feature importance per base estimators group (C). Describe each of the R cases subgroups in terms of features and values. We use Exceptional Model Mining for that task. Performance of the ensemble against the dataset compared to performance of base estimator groups against subgroups of cases, adds transparency.