Self-training field pattern prediction based on kernel methods

Conventional predictors often regard input samples as identically and independently distributed (i.i.d.). Such an assumption does not always hold in many real scenarios, especially when patterns occur as groups, where each group shares a homogeneous style. These tasks are named as the field prediction, which can be divided into the field classification and the field regression. Traditional i.i.d.-based machine learning models would always face degraded performance. By breaking the i.i.d. assump- tion, one novel framework called Field SupportVector Machine (F-SVM) with both classification (F-SVC) and regression (F-SVR) purposes is in- troduced in this chapter. To be specific, the proposed F-SVM predictor is investigated by learning simultaneously both the predictor and the Style Normalization Transformation (SNT) for each group of data (called field). Such joint learning is proved to be even feasible in the high-dimensional kernel space. An efficient alternative optimization algorithm is further designed with the final convergence guaranteed theoretically and experimentally. More importantly, a self-training based kernelized algorithm is also developed to incorporate the F-SVM model with the unknown field during the training phase by learning the transductive SNT to transfer the trained field information to this unknown style data. A series of experiments are conducted to verify the effectiveness of the F-SVM model with both classification and regression tasks by promoting the classification accuracy and declining regression error. Empirical results demonstrate that the proposed F-SVM achieves in several benchmark datasets the best performance so far, significantly better than those state-of-the-art predictors.

Duke Scholars

Author Kaizhu Huang DKU Faculty