Journal of Machine Learning Research () Submitted ; Published
Distributed Primal Coordinate Descent Method for Support Vector Machines
Editor:
Abstract
Keywords:
1. Introduction 2. Experiments
Dataset l n Density Best C
yahoo-japan 176, 203 832, 026 0.016% 0.5 yahoo-korea 460, 554 3, 052, 939 0.011% 2
url 2, 396, 130 3, 231, 961 0.004% 4
webspam 350, 000 16, 609, 143 0.022% 32 KDD2010-a 8, 407, 752 20, 216, 830 0.000% 0.015625 KDD2010-b 19, 264, 097 29, 890, 095 0.000% 0.03125
Table 1: Data statistics. For url, webspam, KDD2010-a, and KDD2010-b, test sets are not available so we randomly split the original data into 80%/20% as training set and test set, respectively.
Acknowledgments
References
c .
(a) yahoo-japan (b) yahoo-korea (c) url
(d) webspam (e) KDD2010-a (f) KDD2010-b
Figure 1: Relative fuction value difference versus trainingg time
(a) yahoo-japan (b) yahoo-korea (c) url
(d) webspam (e) KDD2010-a (f) KDD2010-b
Figure 2: Relative testing accuracy versus trainingg time
2
Distributed Primal Coordinate Descent Method for Support Vector Machines
(a) CDPrimal-A (b) CDPrimal-E (c) TRON
Figure 3: Percentage of computation and communication
3