Discussion and Conclusion
It is arguable that we may have used a too strict stopping condition in DSVM when
C is large. One possibility is to use the stopping tolerance that is proportional to
C. This will reduce the number of iterations so that directly solving linear SVMs
with large C may be possible. However, in Appendix of the technical report
(Chung, Kao, Sun, and Lin 2002), we show that even in these settings, DSVM
with alpha seeding still makes the computational time several times faster than
the original DSVM, especially for large datasets. Moreover, too large stopping
tolerance will cause DSVM stops with wrong solutions.