Default Clustering from Sparse Data Sets

Categorization with a very high missing data rate is seldom studied, especially from a non-probabilistic point of view. This paper proposes a new algorithm called default clustering that relies on default reasoning and uses the local search paradigm. Two kinds of experiments are considered: the first one presents the results obtained on artificial data sets, the second uses an original and real case where political stereotypes are extracted from newspaper articles at the end of the 19th century.