A MARGIN-BASED MODEL WITH A FAST LOCAL SEARCH FOR RULE WEIGHTING AND REDUCTION IN FUZZY RULE-BASED CLASSIFICATION SYSTEMS
Mohammad Taheri, Mansoor Zolghadri Jahromi · 2014
Fuzzy Rule-Based Classication Systems (FRBCS) are highly in- vestigated by researchers due to their noise-stability and interpretability. Un- fortunately, generating a rule-base which is suciently both accurate and inter- pretable, is a hard process. Rule weighting is one of the approaches to improve the accuracy of a pre-generated rule-base without modifying the original rules. Most of the proposed methods by now, may over-t on training data due to generating complex decision boundaries. In this paper, a margin-based opti- mization model is proposed to improve the performance on unseen data. By this model, xed-size margins are dened along the decision boundaries and the rule weights are adjusted such that the marginal space would be empty of training instances as much as possible. This model is proposed to support the single-winner reasoning method with a special cost-function to remove unde- sired eects of noisy instances. The model is proposed to be solved by a fast well-known local search method. With this solving method, a huge amount of irrelevant and redundant rules are removed as a side eect.Two articial and 16 real world datasets from UCI repository are used to show that the proposed method signicantly outperforms other methods with proper choice of the margin size, which is the single parameter of this method.