Classification Based on Possibilistic Likelihood
Résumé
Classification models usually associate one class for each new instance. This kind of prediction doesn’t reflect the uncertainty that is inherent in any machine learning algorithm. Probabilistic approaches rather focus on computing a probability distribution over the classes. However, making such a computation may be tricky and requires a large amount of data. In this paper, we propose a method based on the notion of possibilistic likelihood in order to learn a model that associates a possibility distribution over the classes to a new instance. Possibility distributions are here viewed as an upper bound of a family of probability distributions. This allows us to capture the epistemic uncertainty associated with the model in a faithful way. The model is based on a set of kernel functions and is obtained through an optimization process performed by a particle swarm algorithm. We experiment our method on benchmark dataset and compares it with a naive Bayes classifier.