Clustering is a fundamental task in data analysis, aimed at grouping objects into clusters based on similarity. Traditional methods, like k-means, assign each object to a single cluster, which is not appropriate when uncertainty or overlap between clusters is important. Belief function-based approaches, derived from Dempster-Shafer theory, allow to manage this uncertainty by assigning each object several degrees of membership.

In this thesis, we contribute to the development of frameworks and new algorithms for evidential clustering of multidimensional sequential trajectories (MST). We examined and comparing existing approaches, including feature-based, raw-data-based and model-based methods, focusing on their application to categorical or discrete time series. We explore new frameworks, including the use of traditional evidential c-means (ECM) algorithms combined with feature extraction and unsupervised  feature  selection methods. We proposed new clustering algorithms based on belief functions for MST clustering, such as multi-view evidential c-medoids (MECMdd) and soft evidential c-means (Soft-ECM). The effectiveness of the proposed frameworks and new algorithms is estimated on different synthetics and real datasets. We also applied our methods to chronic pain care trajectories analysis as part of the eDOL project.