Using Huffman Trees in Features Selection to Enhance Performance in Spam Detection

Abstract

Spam detection is very costly when compared to the simple task of spreading spam. Most approaches aim to reach higher accuracy percentages, leaving the classification performance in background, what may cause many problems, such as bottlenecks in the e-mail system, huge infrastructure investments and waste of resources pooling. To avoid these problems, this paper proposes a hierarchical spam features organization using Huffman Trees, where the most important features stay closer to the root. With the reduction of these trees (leaves pruning) the feature space is significantly reduced, speeding up the e-mail classification process. The experiments showed a performance 60 times faster when compared to Spam Assassin.