A distance-based classifier built on class model

In this paper we propose a new type of distance-based classifier. Traditionally, these classifiers are instancebased: they classify a test instance by computation of a similarity measure between that instance and the instances in the training-set and assigning it the same class of the most similar k instances. This method is simple but has some disadvantages, among which there is the greater sensitivity to the local presence of noise in the training-set, the need to compute many similarity measures and the difficulty to decide the value of k. In our approach the classifier is model-based: it first constructs a probabilistic model for each class based on a selection of the itemsets, weighted by their probability to appear in the training data instances of the given class. Then it uses these class models to predict the class of test instances by calculating the distance between the test instance and each of the class model. We have experimented with four different proximity measures and different metrics of itemsets selection on a large collection of datasets from UCI archive. We show that our methods has many benefits. It reduces the number of distance computations. It improves classification accuracy of state-of-the art classifiers, like decision trees, SVM, k-nn, Naive Bayes, rulebased classifiers and association rule-based ones. Finally, it is less sensitive to overfitting and to noise. We perform also an automatic tuning of the algorithmic parameters like the itemsets frequency threshold and the itemsets selection.