Transfer learning via Laplacian feature embedding
Jiangwei Qin, Xiaocun Sun, Jie Zhao, Shizhong Jiang · 4th International Conference on Information Science, Electrical, and Automation Engineering (ISEAE 2022) · 2022
In the real applications, labeled data is often in short supply, which make is difficult to train a reliable classifier. Transfer learning theory enables model to be trained by exploiting the auxiliary data from out domains. For transfer learning, it’s crucial to find a new feature representation where traditional classifier can be applied. In this paper, we propose a novel transfer algorithm by constructing a Laplacian feature embedding where data property is preserved and domain discrepancy is reduced by MMD principle. Furthermore, we exploit the label constraint during the embedding procedure which minimizes the empirical error on the labeled data. Extensive experiments are conducted to evaluate the effectiveness on two textual datasets, and the study shows the proposed method can significantly improve the classification accuracy over several state-of-the-art algorithms.