Abstract: Several experiences proved the impact of the choice of the parts of documents selected on the result of the classification and consequently on the number of requests which can answer these clusters. The process of aggregation gives a very natural method of data classification and considers then m produced classifications by them m attributes and tries to produce a classification called "optimal" which is the most close possible of m classifications. The optimization consists in minimizing the number of pairs of objects (u, v) such as a C classification place them in the same cluster whereas another C' classification place them in different clusters. This number corresponds to the concept of disagreements. We propose an approach which exploits the various elements of an XML document participating in various views to give different classifications. These classifications are then aggregated in the only one classification minimizing the number of disagreements. Our approach is divided into two steps: the first consists in applying the K-means algorithm on the collection of XML documents by considering every time a different element from the document. Second step aggregates the various classifications obtained previously to produce the one that minimizes the number of disagreements.