Compression-based AODE classifiers

Corani, Giorgio and Antonucci, Alessandro and De Rosa, Rocco (2012) Compression-based AODE classifiers. In: 20th European Conference on Artificial Intelligence (ECAI 2012), 27.08.2012-31.08.2012, Montpellier, France.

Compression-based AODE classifiers.pdf - Published Version

Download (241kB) | Preview


The compression-based approach [1] averages over different models, by applying a logarithmic smoothing to the models' posterior probabilities; it has shown excellent performances when applied to ensemble of naive Bayes classifiers. A different ensemble of models is the AODE classifier [11], which is an ensemble of non-naive classifiers (SPODEs), whose probabilistic predictions are ag-gregated by a simple arithmetic mean; it achieves high performance, often outperforming also more sophisticated aggregation schemes over SPODEs. Our first contribution is COMP-AODE, which applies the compression-based approach over SPODEs, getting an overall better classification performance than AODE. Then, we substitute the uniform prior over SPODEs by a set of prior mass functions, developing the credal classifier COMP-AODE*. The main feature of COMP-AODE* is that it returns more classes instead of a single one when the classification is prior-dependent, namely if the most probable class varies when different priors over the SPODEs are considered. Extensive experiments show that, thanks to the extension to imprecise probability, COMP-AODE* offers better classification performance than COMP-AODE.

Actions (login required)

View Item View Item