Bitcoin is the most used and studied cryptocurrency. Being decentralized, transaction data is freely accessible and can thus be analyzed. The first step to most analyses consists in grouping individual – anonymous – addresses in clusters, corresponding to Bitcoin actors. In this article, we propose a new method to realize these machine learning-based aggregates. Our approach is based on the construction of a training dataset whose class variable is obtained by a ground truth computed a posteriori. This dataset is used to identify the change addresses of transactions, addresses belonging to the author of the transaction. This makes it possible to increase the number of addresses discovered as belonging to the same actor. Using an external validation criterion, we experimentally show the relevance of this method in comparison with the heuristics usually used.