next up previous contents
Next: 5.2.1 Indirect (IRLS) Stability Up: 5. Logistic Regression Computation Previous: 5.1.6 Scope   Contents


5.2 IRLS Parameter Evaluation and Elimination


Table 5.3: IRLS Parameters
Parameter Description
modelmin Lower threshold for $ \mu = \exp(\eta) / (1+\exp(\eta))$
modelmax Upper threshold for $ \mu = \exp(\eta) / (1+\exp(\eta))$
wmargin Symmetric threshold for weights $ w = \mu (1-\mu)$
margin Symmetric threshold for outcomes $ y$
binitmean Initialize $ \beta_0$ to E( $ \mathbf{y}$)
rrlambda Ridge-regression parameter $ \lambda$
cgwindow Number of non-improving iterations allowed
cgdecay Factor by which deviance may decay during iterations
lreps Deviance epsilon for IRLS iterations
cgeps Residual epsilon for CG iterations
cgdeveps Deviance epsilon for CG iterations
lrmax Maximum number of IRLS iterations
cgmax Maximum number of CG iterations
cgbinit CG iterations for IRLS iteration $ i$ start where IRLS iteration $ i-1$ stopped

Figure 5.3: LR tree with rectangle marking the beginning of our IRLS implementation.
\includegraphics{figures/treemap-methods-IRLS.eps}



Subsections
next up previous contents
Next: 5.2.1 Indirect (IRLS) Stability Up: 5. Logistic Regression Computation Previous: 5.1.6 Scope   Contents
Copyright 2004 Paul Komarek, komarek@cmu.edu