Matching at large scale represents a laborious process. The standard approach trying to match the complete input schemas will often lead to performance problems matching. In this context, we propose a pre-matching approach based on a XML schemas reduction. The presented approach identifies common structures between and within XML schemas and attempts to scale down these input schemas. Our method uses tree mining techniques to identify these common structures. The interest of this approach is to reduce the size of the XML schemas before matching and to find, at the same time, the similarity of sub-parts of the schemas. As shown by our experiments, resulting schemas are much smaller than originals one which will improve matching performance.