One of the simplest methods for two-group classification is naïve Bayes, in which predictors are treated as though they provide independent information. Traditional event models underlying naive Bayes classifiers assume probability distributions that are not appropriate for binary data generated by human behavior. This paper develops a new event model, based on a somewhat forgotten distribution created by K.T. Wallenius more than 50 years ago. The authors show that it achieves superior performance using less data on a collection of Facebook datasets, where the task is to predict personality traits, based on likes.
Read the paper:
Wallenius Naïve Bayes. E.J. de Fortuny, D. Martens, F. Provost.