Abstract: Semi-supervised learning is a hot research topic of machine learning. Co-training is a multi-view semi-supervised learning method. Uraph representation was introduced to co-training where multiple graphs were used to represent multi-relation data. Semi-supervised learning was processed on each graph while co-training was conducted between graphs to ensure the predictions of graphs arc the same. A new co-training algorithm for multi-relation data was proposed,and it was analyzed from the viewpoint of probability. Encouraging experimental results are gotten from real world multi-relation dataset.

Key words: Machine learning, Semi-supervised learning, Co-training, Multi-relation data