Vers une approche utilisant l'apprentissage de métrique pour du clustering semi-supervisé interactif d'images

The problem of unsupervised and semi-supervised clustering is extensively studied in machine learning. In order to involve user in image data clustering, (Lai et al., 2014) proposed a new approache for interactive semi-supervised clustering that translates the feedback of user (expressed at the level of individual images) into pairwise constraints between groups of images, these groupes being constituted thanks to the underlying hierarchical clustering solution and user feedback. Recently, the need for appropriate measures of distance or similarity between data led to the emergence of distance metric learning approaches. In this paper, we propose a method incorporating metric learning in the existing system of (Lai et al., 2014) to improve performance and reduce the computational time. Our experimental results obtained on the image database Wang show that the metric learning methods fit well into the existing system improve its performances as well as its computational time. MOTS-CLÉS : clustering semi-supervisé interactif, contraintes par paires, apprentissage de métrique, analyse d’images.