Weighted Multi-source TrAdaBoost

In order to take full advantage of valuable information from all source domains and to avoid negative transfer resulted from irrelevant information, a kind of weighted multi-source TrAdaBoost algorithm is proposed. At first, some weak classifiers are respectively trained based on training sample sets constituted by both each source domain and the target domain. Then we assign a weight to each weak classifier according to its error on the target training set. In the third step, a candidate classifier is obtained based on the weighted sum of all weak classifiers. In the fourth step, sample weights of the source and target domains are updated according to the error of the candidate classifier on corresponding domains. At last, all weak classifiers are retrained based on the training samples with new updated weights. The above steps repeated until the number of maximum iterations is reached. Experimental results on bimonthly datasets show that, compared with TrAdaBoost and multi-source TrAdaBoost, the proposed algorithm has higher classification accuracy.