Cost-Sensitive Ensemble via Adaptive Weighted Cost Proportionate Sampling
Xin Luo -, Qingsheng Zhu · International Journal of Digital Content Technology and its Applications · 2011
Cost Proportionate Sampling (CPS) has proved to be a feasible and highly efficient cost-sensitive meta-learning technique, however, current cost-sensitive meta-learning techniques based on CPS suffer the problem of stationary sampling probability which will decrease the diversity between samples and result in low efficiency for ensemble. This paper presents a novel CPS technique named Adaptive Weighted Cost Proportionate Sampling (AWCPS), which works by introducing an adaptive selection weight as a balance factor during CPS process to diversify the samples as well as maintaining the cost-sensitivity. Based on AWCPS, we further propose a novel cost-sensitive metalearning algorithm named Cost Sensitive Ensemble via AWCPS (CSEA). The experiments on public large, real datasets demonstrate that compared to current cost-sensitive algorithms based on CPS, CSEA can obtain significant advantage in degrading the cumulative costs as well as improve the classification accuracy.