An Enhanced TranCo-Training Categorization Model with Transfer Learning
Huan Tang · 2013
When unlabeled data draw from different distributions compared with labeled data in semi-superviselearning, the topic biases the target domain and the performance of semi-supervised classifier decreases.The transfer technique is applied to improve the performance of semi-supervised learning in this paper.An enhanced categorization model called TranCo-training is studied which combines transfer learningtechniques with co-training methods. The transferability of each unlabeled instance is computed by animportant component of TranCo-training according to the consistency with its labeled neighbors.At eachiteration, unlabeled instances are transferred from auxiliary dataset according to their transfer ability.Theoretical analysis indicates that transfer ability of an unlabeled instance is inversely proportional to itstraining error, which minimizes the training error and avoids negative transfer.Thereby, the problem oftopic bias in semi-supervised learning is solved. The experimental results show that TranCo-training algorithm achieves better performance than the RdCo-training algorithm when a few labeled data on targetdomain and abundant unlabeled data on auxiliary domain are provided.