Une nouvelle métrique pour l'analyse discriminante sur données de dissimilitude

Statistical pattern recognition traditionally relies on a features based representation. For many applications, such vector representation is not available and we have only proximity data (distance, dissimilarity, similarity, ranks ...). In this paper, we consider a particular point of view on discriminant analysis from dissimilarity data. Our approach is inspired by the Mahalanobis distance. We define decision rules to mimic the behaviour of a linear or a quadratic Gaussian classifier. The number of parameter is limited (two per class). Numerical experiments on real data show an interesting behaviour compared to a KNN classifier (i) lower or equivalent error rate, (ii) better robustness with sparse dissimilarity data.