Plongement incrémental dans un contexte de dissimilarité
Résumé
Statistical pattern recognition framework is based on a numerical description of objects and can thus be easily combined with efficient machine learning methods. On the other hand structural pattern recognition methods use a limited set of machine learning methods but encode a rich description of objects through structural models such as strings or graphs. This last decade have seen the emergence of two closely related trends aiming at bridging the gap between these two frameworks by combining their respective advantages: Graph or string ker- nels in one hand and dissimilarity representation on the other hand. However, an important family of dissimilarity representation methods requires the whole universe to be known during the learning phase in order to build an explicit embedding of structural data which can then be combined with any machine learning methods. This latter property is an important limitation in many practical applications where the test set is unbounded and unknown during the learn- ing phase. Moreover requiring the whole universe represents a bottleneck for the processing of massive dataset. We propose in this paper to overcome this last limitation and show the connection of this solution with the kernel framework.