next up previous contents
Next: 5.2.1.8 Basic Stability Test: Up: 5.2.1 Indirect (IRLS) Stability Previous: 5.2.1.6 Basic Stability Test:   Contents


5.2.1.7 Basic Stability Test: Sparse Conclusions

This concludes our basic stability tests with the modelmin, modelmax, margin, rrlambda, cgwindow and cgdecay parameters on sparse datasets. The most obvious conclusions are that modelmin, modelmax and margin have little beneficial effect, and sometimes are detrimental. It appears that rrlambda is very effective in preventing over-fitting, and in avoiding the wasted calculations and seconds that accompany over-fitting. The cgwindow and cgdecay parameters are especially helpful when rrlambda is not used, and may also be helpful when rrlambda is used. In no case were rrlambda, cgwindow or cgdecay particularly detrimental. We have raised the question as to whether multiplying cgeps by the number of attributes is appropriate. Perhaps more sensible would be multiplying cgeps by a different factor, as discussed later in Section 5.2.2.


next up previous contents
Next: 5.2.1.8 Basic Stability Test: Up: 5.2.1 Indirect (IRLS) Stability Previous: 5.2.1.6 Basic Stability Test:   Contents
Copyright 2004 Paul Komarek, komarek@cmu.edu