Three-Phase Tournament-Based Method for Better Email Classification

Email classification performance has attracted much attention in the last decades. This paper proposes atournament-based method to evolve email classification performance utilizing World Final Cup rules as asolution heuristics. Our proposed classification method passes through three phases: 1) clustering(grouping) email folders (topics or classes) based on their token and field similarities, 2) training binaryclassifiers on each class pair and 3) applying 2-layer tournament method for the classifiers of the relatedclasses in the resultant clusters. The first phase evolves K-mean algorithm to result in cluster sizes of 3, 4,or 5 email classes with the pairwise similarity function. The second phase uses two classifiers namelyMaximum Entropy (MaxEnt) and Winnow. The third phase uses a 2-layer tournament method whichapplies round robin and elimination tournament methods sequentially to realize the winner class percluster and the winner of all clusters respectively. The proposed method is tested for various K settingsagainst tournament and N-way methods using 10-fold cross-validation evaluation method on Enronbenchmark dataset. The experiments prove that the proposed method is generally more accurate than theothers.