Learning-Based Spam Filters: the Influence of the Temporal Distribution of Training Data

The great number and variety of learning-based spam filters proposed during the last years cause the need in complex and many-sided evaluation of them, taking features of the phenomenon of spam into account. This paper is dedicated to the analysis of the dependence of filter performance on the temporal distribution of training data; the cause of this dependence is the changeability of email. Such analysis provides additional information about the filter quality, and also may be useful for organizing more effective training of the filter. The native Bayes filter is chosen for evaluation in this paper.