Abstract:

Multi-label learning usually has many unlabeled samples. Combined with co-training method, this research made full use of the unlabeled sampled in dataset, selected the local k-NN(k nearest neighbor) and global k-NN for training to get two classifiers, which could label the unlabeled examples and could be added to the training set. The collaborative training process iterated continuously, until the training finished. The experimental results showed that this algorithm could outperform other multi-label learning algorithms.

Key words: multi-label learning, global k-NN, semi-supervised learning, local k-NN