Speaker: Eric Gaussier, Director of the LIG Computer Science Laboratory in Grenoble
Title Learning representations for clustering: some solutions to all-posed problem.
Abstract

Clustering aims at grouping similar objects in the same clusters. The problem is however partly subjective, and hence ill-defined, as, for a given collection of documents, different users may be interested in different clusters. We propose to solve this problem by learning representations of documents that are better adapted to the clustering problem and the user needs. To do so, we rely on deep clustering that consists in identifying clusters in an embedded space obtained via autoencoders.