In　order　to　learn　ensembled　function　and　improve　classification　performance,　a　new　algorithm　framework　named　TPEL　(Two-Phases　Ensemble　Learning)　is　proposed.　For　the　task　of　email　filtering,　a　typical　problem　of　two-class　categorization,　we　conduct　a　series　of　experiments　on　four　public　available　datasets.　The　experimental　results　show　that　firstly　the　performance　of　TPEL　is　faintly　affected　by　the　count　of　the　combined　classifiers.　Secondly,　TPEL　bears　the　best　capacity　when　it　combines　multiple　heterogeneous　classifiers.　Thirdly,　in　most　of　the　experiments,　the　performance　of　TPEL　is　better　than　that　of　the　comparing　algorithms　such　as　NalIve　Bayes,　Bagging,　Boosting　etc.　In　addition,　TPEL　reveals　its　promising　results　in　the　situation　of　that　either　the　weak　learner　is　steady　or　not.　At　last,　TPEL　is　provided　with　reasonable　time　complexity.