In fuzzy rule-based classification systems, a large number of descriptive attributes leads to an explosion of the generated rules' number and may affect the accuracy of learning algorithms. In order to address this problem, a solution is to treat separately subgroups of attributes. This allows decomposing the learning problem into sub-problems of lower complexity, and getting more intelligible rules as they are smaller. We propose a new method to regroup attributes; it is based on the concept of association rules. These rules highlight interesting relationships between value ranges of attributes. These local associations are then aggregated at the attributes' level according to the number of found associations and to their significance. Our approach, tested on different learning bases and compared to the classical approach (SIF), allows improving the accuracy and guarantees a reduction of the rules' number.