next up previous contents
Next: 5.2.1.11 Stability Test: wmargin Up: 5.2.1 Indirect (IRLS) Stability Previous: 5.2.1.9 Basic Stability Test:   Contents


5.2.1.10 Basic Stability Test: Dense Conclusions

Our main conclusion for dense datasets is that stability is less of an issue than for our sparse datasets. One reasonable hypothesis is that stability problems are correlated with data having a large number of nonzero attributes in some rows. Having many nonzero values in a single record may lead to some rows exceeding the numerical bounds described at the beginning of Section 5.2.1.1. Another hypothesis, overlapping the first, is that datasets with a widely varying number of nonzero attributes per row are more susceptible to stability problems. Such a dataset could easily cause the entries in the $ \ensuremath{\mathbf{X}}^T \ensuremath{\mathbf{W}} \ensuremath{\mathbf{X}}$ matrix to be badly scaled. In any case, we believe our data suggests that modelmin, modelmax, margin and rrlambda are unnecessary for dense data, while cgwindow and cgdecay may be harmful or at least require careful attention.


next up previous contents
Next: 5.2.1.11 Stability Test: wmargin Up: 5.2.1 Indirect (IRLS) Stability Previous: 5.2.1.9 Basic Stability Test:   Contents
Copyright 2004 Paul Komarek, komarek@cmu.edu