Improving Explanatory Power of Machine Learning in the Symbolic Data Analysis Framework

摘要

Many nice machine learning methods are black box producing very efficient rules but hard to be understandable by the users. The aim of this paper is to help user by tools allowing a better comprehension of these rules. These tools are based on characteristic properties of the original variables in order to remain in the natural language of the user. They are based on three principles, first on local models fitting at best clusters to be found, second on a symbolic description of these clusters and their Symbolic Data Analysis, third on characteristic criterion increasing the explanatory power of the rules by an adaptive process filtering explanatory sub populations.