A Hybrid, Case-Based Related Approach to Generate Predictions from Rules
Résumé
This work takes place in the general context of the construction of a prediction for decision support issues. It relies on knowledge expressed by numerous rules with homogeneous structure, extracted from various scientific publications of a domain. In this paper we propose a predictive approach that allows one to perform two stages: firstly, the generation of a partition of the rules into groups that express a common experimental tendency; secondly, the computation of a prediction rule, starting from a new description of experimental conditions and from the obtained groups of rules.The method is experimented on a case study in food science. Compared to the results that are obtained by a classical approach based on a decision tree classifier, the proposed method obtains good predictions, in the sense of accuracy, completeness and error rate.