“BOF” Trees Diagram as a Visual Way to Improve Interpretability of Tree Ensembles

Abstract

The motivation for this research stemmed from a desire to create visual aids to help researchers/managers interpret ensembles of decision tree outputs generated by various algorithms. The method employed a simulation experiment (using only bagging) followed by application of the new visualization tools on actual survey data. Simulated data, with a pre-specified structure, were "bagged" with the results presented using five graphical tools that recreated (and/or portrayed) the known data structures captured by the bagging algorithm. Then the same methodology was generalized to a structurally unknown, virgin (survey) data set. Results of the research are that five visual aids tools were examined (two of which are new approaches) and found to be useful for making action-oriented interpretations from e.g., web-survey data.