In the wish list of the characteristics of a classifier, there are a reliable approach to small data sets and a clear and robust treatment of incomplete samples. This paper copes with such difficult problems by adopting the paradigm of credal classification. By exploiting Walley's imprecise Dirichlet model, it defines how to infer the naive credal classifier from a possibly incomplete multinomial sample. The derived procedure is exact and linear in the number of attributes. The obtained classifier is robust to small data sets and to all the possible missingness mechanisms. The results of some experimental analyses that compare the naive credal classifier with naive Bayesian models support the presented approach.
Keywords. credal classification, classification, naive credal classifier, naive Bayes classifier, credal sets, imprecise Dirichlet model, inference, missing data, incomplete samples
Format. PDF
Paper Download
The paper is availabe in the following sites:
Authors addresses:
Galleria 2
CH-6928 Manno
Switzerland
E-mail addresses:
| Marco Zaffalon | zaffalon@idsia.ch |