Decision tree induction methods using an entropy criterion – II. Local approaches
Résumé
A great number of systems can only be described by models established through the statistical analysis of empirical data. When both the output and the input data are qualitative, automatic learning algorithms have to be designed. Those algorithms generate rules which describe the input/output behaviour of the system, from the observed data. In this contribution, we propose to generate non deterministic rules from a set of incoherent data. We first introduce an information based index in order to measure the incoherence of the learning set (the feasibility of the modelling problem is thus measured). Then, we propose a procedure which allows to increase the coherence of the learning set, by considering only a subset for which the feasibility index is greater. The determination of such a subset is made in an optimal way, such that the lost information is minimized. In order to construct simpler models, we present five local (or « by node ») approaches based on the construction of a decision tree. These methods are characterized by a top-down construction algorithm, even if, for each node, we determine the variable to test with a bottom-up algorithm in the case of the mixt approach. Taking into account the best by level method developed in [Per97], we apply the presented methods on several databases, in order to compare them.