Using GNUsmail to Compare Data Stream Mining Methods for On-line Email Classification

Real-time classification of emails is a challenging task because of its online nature, and also because email streams are subject to concept drift. Identifying email spam, where only two different labels or classes are defined (spam or not spam), has received great attention in the literature. We are nevertheless interested in a more specific classification where multiple folders exist, which is an additional source of complexity: the class can have a very large number of different values. Moreover, neither cross-validation nor other sampling procedures are suitable for evaluation in data stream contexts, which is why other metrics, like the prequential error, have been proposed. However, the prequential error poses some problems, which can be alleviated by using recently proposed mechanisms such as fading factors. In this paper, we present GNUsmail, an open-source extensible framework for email classification, and we focus on its ability to perform online evaluation. GNUsmails architecture supports incremental and online learning, and it can be used to compare different data stream mining methods, using state-of-art online evaluation metrics. Besides describing the framework, characterized by two overlapping phases, we show how it can be used to compare different algorithms in order to find the most appropriate one. The GNUsmail source code includes a tool for launching replicable experiments.