Clustering et intégration de connaissances
Nguyen-Viet-Dung Nghiem · theses.fr (ABES) · 2021
Clustering is one of the essential topics in data mining. Although it is designed to work in a fully unsupervisedway, its application in real-world data is often regulated by expert knowledge. Constrained clustering (ageneralization of semi-supervised clustering) aims to exploit this knowledge during the clustering task. In thisthesis, we develop two frameworks to integrate expert constraints in the clustering task. In the first work, wepropose a declarative post-processing method to adapt the output of a clustering algorithm to satisfy theconstraints. The originality is to consider an allocation matrix that gives the scores for attribution of points toeach cluster and to find the best partition satisfying all the constraints. In the second work, we propose aunified framework to integrate general constraints in a clustering model with deep learning. The genericity isobtained by formulating the constraints in propositional logic, defining two versions of semantic loss, andcomputing them through Weighted Model Counting. Experimental results on well-known datasets show thatour approach is competitive with other constraint-specific methods while being general. In addition, we havedefined and formulated new types of constraints in clustering: the cluster coverage constraint limiting thenumber of clusters to which a group of points can belong and the combined fairness constraint taking intoaccount both the group fairness and individual fairness.