35 rue Saint-Honoré, Fontainebleau

The graph matching problem is among the most important challenges of graph processing, and plays a central role in various fields of pattern recognition. We propose an approximate method for labeled weighted graph matching, based on a convexconcave programming approach which can be applied to the matching of large sized graphs. This method allows to easily integrate information on graph label similarities into the optimization problem, and therefore to perform labeled weighted graph matching. One of the interesting applications of the graph matching problem is the alignment of protein-protein interaction networks. This problem is important when investigating evolutionary conserved pathways or protein complexes across species, and to help in the identification of functional orthologs through the detection of conserved interactions. We reformulate PPI alignment as a graph matching problem, and study how state-of-the-art graph matching algorithms can be used for this purpose. In the classical formulation of graph matching, only one-to-one correspondences are considered, which is not always appropriate. In many applications, it is more interesting to consider many-to-many correspondences between graph vertices. We propose a reformulation of the many-to-many graph matching problem as a discrete optimization problem and we propose an approximate algorithm based on a continuous relaxation. In this thesis, we also present two interesting results in statistical machine translation and bioinformatics. We show how the phrase-based statistical machine translation decoding problem can be reformulated as a Traveling Salesman Problem. We also propose a new protein binding pocket similarity measure based on a comparison of 3D atom clouds.