Transfer Learning on High Variety Domains for Activity Recognition
Josh Jia-Ching Ying, Bo-Hau Lin, Vincent S. M. Tseng, Sun‐Yuan Hsieh · 2015
The research topic on transfer learning task has attracted a lot of attentions in recent years due to the wide applications. Although a number of transfer learning techniques have been developed, basically they were designed in the manner of learning and transferring among multiple source domains and it was assumed that the source domains and target domain share the same feature space. However, with the high variety issue under big data environments, this assumption violates the scenario of many real-world applications like activity recognition. In this paper, we propose a novel approach for transfer learning on activity recognition with the new concept of transfer learning on high variety domains. The core idea of our transferring model is based on theoretical statistic hypothesis tests, Kolmogorov-Smirnov test and x2 goodness of fit test, which evaluate how well a domain is covered by another domain based on similarity between each pair of features. Through comprehensive evaluations by experiments, our proposal is shown to deliver excellent effectiveness and substantially outperform state-of-the-art multiple source domain transfer learning methods. To our best knowledge, this is the first work that explores the problem of transfer learning on high variety domains for activity recognition with promising potential in wide applications.