When working with high-dimensional data, visualization techniques are useful tools to help us discover patterns. However, the visualization results do not always match the users’ expectations. An effective mechanism that allows human users to interact with machine learning systems is thus necessary. This thesis addresses the problem of combining users’ feedback and dimensionality reduction (DR) techniques for visualization. From a machine learning viewpoint, users’ feedback is considered as constraints. Our goal is to transform users’ constraints into adequate terms that can enrich DR methods and assess the visualization. We propose novel interactive visualization methods for integrating semantic information or users’ prior knowledge directly into the visualization. New usage of constraints for evaluating the quality of a visualization is also proposed. This thesis focuses not only on the representation of different kinds of users’ constraints but also on the computational efficiency of each proposed method.
|la date de réponse||1 déc. 2021|
|Superviseur||Benoît Frénay (Promoteur), Adrien Bibal (Copromoteur), Wim VANHOOF (Président), Bruno Dumas (Jury), John Aldo Lee (Jury) & Michel Verleysen (Jury)|