Fuzzy Domain Adaptation Using Unlabeled Target Data
Transfer learning has been emerging recently and gaining more attention because of its ability to deal with “small labeled data” issue in new markets and for new products. It addresses the problem of leveraging knowledge acquired from previous domain (a source domain with a large amount of labeled data) to improve the accuracy of tasks in the current domain (a target domain with little labeled data). Fuzzy rule-based transfer learning methods are developed due to the ability to dealing with the uncertainty in domain adaptation scenarios. Although some effort is made to develop the fuzzy methods, they only apply the knowledge of the labeled data in the target domain to assist the model’s construction. This work develops a new method that explores and utilizes the information contained in the unlabeled target data to improve the performance of the new constructed model. The experiments on both synthetic datasets and real-world datasets illustrate the effectiveness of our method, and also give the application scope of applying it.
KeywordsDomain adaptation Transfer learning Machine learning Fuzzy rules Regression
This work was supported by the Australian Research Council under DP 170101623.
- 7.Gönen, M., Margolin, A.A.: Kernelized Bayesian transfer learning. In: AAAI, pp. 1831–1839 (2014)Google Scholar
- 9.Bengio, Y.: Deep learning of representations for unsupervised and transfer learning. In Proceedings of ICML Workshop on Unsupervised and Transfer Learning, pp. 17–36 (2012)Google Scholar
- 14.Rasmussen, C.E.: The infinite Gaussian mixture model. In: Advances in Neural Information Processing Systems, pp. 554–560 (2000)Google Scholar