Term Space Partition Based Ensemble Feature Construction for Spam Detection

This paper proposes an ensemble feature construction method for spam detection by using the term space partition (TSP) approach, which aims to establish a mechanism to make terms play more sufficient and rational roles by dividing the original term space and constructing discriminative features on distinct subspaces. The ensemble features are constructed by taking both global and local features of emails into account in feature perspective, where variable-length sliding window technique is adopted. Experiments conducted on five benchmark corpora suggest that the ensemble feature construction method far outperforms not only the traditional and most widely used bag-of-words model, but also the heuristic and state-of-the-art immune concentration based feature construction approaches. Compared to the original TSP approach, the ensemble method achieves better performance and robustness, providing an alternative mechanism of reliability for different application scenarios.