Self-paced and Bayes-decision-rule linear KNN prediction

While a testing sample may be first encoded linearly with labeled samples and then classified with KNN on the sum of the obtained weights of the samples in each class so as to avoid the consistent distribution assumption explicitly or implicitly used in most of the existing classification methods for training and testing samples, a novel self-paced and Bayes-decision-rule linear KNN prediction method SBLD-KNN in this study will be proposed to achieve three goals: (1) class-ware information will be explicitly reflected in a grouping effect regularization term so as to share the sparsity of a linear encoder and simultaneously have grouping effect of weights on each class; (2) the resultant predictor behaves like Bayes-decision-rule for minimum error; (3) self-paced regularized term is designed to adaptively truncate the weights of labeled samples for enhancing generalization. In order to do so, the corresponding objective function of SBLD-KNN is designed and then optimized by using the alternating optimization strategy, and its Bayes-decision-rule is theoretically analyzed. Our experimental results on benchmark datasets witness the effectiveness of SBLD-KNN, in contrast to the comparative methods, including SBLD-KNN’s simplified version BD-KNN with weight’s truncating rather than self-pacing.