Online Laplacian-Regularized Support Vector Regression
Lianbo Zhang, Weifeng Liu · 2017
In recent years, with the growing quantity of data and the explosion in the amount of available information, much effort has been devoted by researchers to develop better learning methods to address these issues. Among these methods, online learning is one set of instrumental methods that could fit such learning scenarios through training one example at a time. Specifically, online support vector regression (SVR) is a typical online learning method that has been developed on the basis of support vector machines (SVM) and is able to achieve effective performance. However, it fails to adequately utilize a large mass of unlabelled data. Meanwhile, it has been shown that manifold learning enables us to exploit the geometry of the data distribution, but it remains a challenge to explore the online manifold process. In this paper, we propose a new online method that incorporates standard online SVR in a manifold regularized framework, which we call online Laplacian-regularized SVR (online LapSVR). Our online LapSVR algorithm can tackle the problems described above and avoid the repeated calculations of batch methods, which require intensive computation, especially with data of massive size or dimensions. Experimental evidence is presented for the Housing, Auto-MPG and DEAP datasets that suggests that our algorithm is able to perform better than standard online SVR methods.