This　paper　proposes　a　novel　ensemble　learning　framework,　namely,　Multiple-Phase　Cost-Sensitive　Ensemble　Learning　(MPCSL)　which　simulates　the　means　and　process　of　human　being　learning.　It　consists　of　two　types　learning,　i.e.,　direct　learning　which　learns　multiple　weak　learners　from　a　training　dataset　via　some　homogeneous　or　heterogenous　algorithms,　and　indirect　learning　that　constructs　a　committee　from　the　knowledge　of　the　combined　filters　or　other　committees.　This　paper　studies　empirically　the　performance　of　MPCSL　on　spam　filtering　tasks.　In　the　occasions　of　combining　homogeneous　and　heterogenous,　how　the　performance　of　MPCSL　changes　is　surveyed.　The　results　shows　that　MPCSL　is　a　compellent　ensemble　learning　method　for　cost-sensitive　tasks　such　as　spam　filtering.