Ensemble Transfer Learning Model Based On Distribution Similarity

Transfer learning is a type of machine learning methods for situations where the source and target domains are distributed differently and it has been extensively studied by scholars.We propose a new ensemble learning model based on distribution similarity which aims to make full use of environmental information by fully focusing on the data in the most similar environments in the source domain during test stage.We use the framework of instance weighting self-adaption and propose a new instance weighting estimation.To compare our method with global method,we set up simulated numerical experiments in various data scenarios.The experiments show that our method works better than training the learner directly on global data in multiple data scenarios.We also conduct data analysis experiments on various configurations of our method to study the preferences of the method.Based on our method we propose the resisting irrelevant variables strategy and the transfer alert strategy.The resisting-irrelevant-variables strategy can be used to improve transfer effectiveness when there are some irrelevant variables in data.The transfer alert strategy can be used to determine whether our method should be applied in a specific problem in advance.We conduct simulated numerical experiments to study the effectiveness of transfer alert strategy.Finally,our method is applied to breast cancer recurrence prediction to elaborate its deployment and effectiveness.