Linear regression (LR) provides a simple and effective approach to address regression problems. To scale up its performance, a number of improved algorithms, such as locally weighted linear regression (LWLR) and the model tree regression algorithms (M5 and its improved algorithm M5P), have been presented. LWLR is a typical lazy learning algorithm. Its high time complexity limits its application. Model trees don't incur the high time complexity confronting LWLR. However, it suffers from tree learning. In this paper, we single out another improved linear regression algorithm called instance weighted linear regression (IWLR) via eager instance weighting. We experimentally tested IWLR using the whole 36 regression datasets selected by Weka, and compared it to LR, LWLR, and M5P. The experimental results show that IWLR significantly outperforms LR and almost ties LWLR and M5P in term of relative mean absolute error. In terms of running time, IWLR is only slower than LR and much faster than LWLR and M5P.
Author's Name: Li, C.
Volume: Volume 4
Issues: Issue 6
Keywords: Instance weighted linear regression, Linear regression, Locally weighted linear regression, Model tree