A Noise-Probability Based Multiple Kernel Boosting Algorithm

To solve the problem that the Multiple Kernel Boosting (MKBoost) algorithm was sensitive to noise, the characteristic that noise data and normal data were treated differently by the optimal classifier was considered, and a MKBoost algorithm applied to datasets polluted by noise data was proposed.The fuse of KNN (K-nearest neighbor) algorithm and logistic regression was used for the construction of noise probability function, and calculating the probability of each instance being noise.A new loss function was constructed via the noise probability function, and addictive model was utilized for computing the coefficients corresponded to base classifier in each iteration.Experiments on the datasets from UCI show that the proposed method could efficiently reduce the sensitivity of MKBoost algorithm to noise, and improved the robustness