The property of χ201 - concordance for Bayesian confirmation measures
[ 1 ] Instytut Informatyki, Wydział Informatyki, Politechnika Poznańska | [ P ] employee
2013
chapter in monograph
english
- interestingness measures
- Bayesian confirmation
- statistical dependency
EN The paper considers evaluation of rules with particular interestingness measures being Bayesian confirmation measures. It analyses the measures with regard to their agreement with a statistically significant dependency between the evidence and the hypothesis. As it turns out, many popular confirmation measures were not defined to possess such a form of agreement. As a result, even in situations when there is only a weak dependency in data, measures could indicate strong confirmation (or disconfirmation), encouraging the user to take some unjustified actions. The paper employs a χ 2-based coefficient allowing to assess the level of dependency between the evidence and hypothesis in experimental data. A method of quantifying the level of agreement (concordance) between this coefficient and the measure being analysed is introduced. Experimental results for 12 popular confirmation measures are additionally visualised with scatter-plots and histograms.
226 - 236