A Study Of Transfer Learning Based On CCA And Co-Training

Traditional machine learning usually assumed that the distributions of training and testing data are consistent, and the training data is sufficient. Since machine learning has been widely used, the distributions of training and testing data are different in many problems, or the training data available are insufficient to train a good classifier. Transfer learning is one way to solve this problem, and the research in this area is of high academic value.As transfer learning significantly improved the predictive performance, it has been widely studied and applied. But researchers found that the algorithm’s weight update policy might lead to a general weight drift problem of training sample, and the algorithm requires a higher similarity between the distributions of the source and target domain samples. This paper presents a new weight update policy. In each training process, the algorithm takes advantage of errors of classification model on both the source and target domain to update the weights of training samples, and then use the newly weighted samples into the next iteration for training. Finally, the algorthm gets a good classifier via a boosting method. The algorithm alleviates the weight drift problem, and weakens the distribution similarity requirements of the source and target domain.In many practical problems, one object can be described in a number of different ways or from different angles. These different descriptions contribute different views of it. Study on relations among views to help transfer learning is of significant research value. This paper uses canonical correlation analysis to process training data, and uses feature fusion to construct a new view for transfer learning. The given algorithm efficiently utilizes the relationship between multi-view data, and produces improving performance.Co-training is a classic method in the problem of multi- view learning. Combination of co-training and transfer learning is of theoretical significance and research value. This paper uses the idea of co-training. In each round of training in transfer leaerning, the algorithm updates the training samples’ weights on two views collaboratively, and then conducts the next round of training. It finally gets a good classification model with the boosting learning method. The multi- view transfer learning algorithm based on co-training exploits the collaborative learning idea to improve the overall performance of the classifier.