Fuzzy Meta-Learning: Preliminary Results

Learning from distributed data is becoming in our times a necessity, but it is also a complex and challenging task. Approaches developed so far have not dealt with the uncertainty, imprecision and vagueness involved in distributed learning. Meta-Learning, a successful approach for distributed data mining, is in this paper extended to handle the imprecision and uncertainty of the local models and the vagueness that characterizes the meta-learning process. The proposed approach, Fuzzy Meta-Learning uses a fuzzy inductive algorithm to meta-learn a global model from the degrees of certainty of the output of local classifiers. This way more accurate models of collective knowledge can be acquired from data with application both to inherently distributed databases and parts of a very large database. Preliminary results are promising and encourage further research towards this direction.