3 ms·
Thesis. It's a proposed classifier algorithm that is suppose to be better than Random Forest and XGBoost for classifying high dimensional data. The data sets a
by anthony_doan 7y ago
Thesis.
It's a proposed classifier algorithm that is suppose to be better than Random Forest and XGBoost for classifying high dimensional data. The data sets are cancer data (prostate and myeloma). Unfortunately it's not going to be publish because I'd like to graduate sooner and that the software does not meet certain criteria for the journal we were aiming for.
The proposed algorithm uses two technique:
1. My forest consist of GUIDE decision trees by Dr. Loh. It is better than CART and M4.5 and such because it does not have the selection bias problem. CART and M4.5 are bias to selecting categorical predictor for node splitting. They're also bias on variables that enable more splitting so decision trees usually contain more levels. GUIDE is also aim at finding interactions candidate to split if it is statistically significant.
2. CERP by Dr. Moon. It makes the trees within the forest less correlated among each other. Much more so than Random Forest. Accuracy takes a hit as your correlation gets higher (obviously zero is the best). It also enabled ensemble of ensembles (ensemble of forests). Of course some of you may state that you can do ensemble of random forests but it is naive and won't help you.
I should be defending in this month or next month.
- splittingTimes 7y agoIs there a public link to your thesis? Can your approach be used to do feature segmentation of 3D (triangle mesh) data?
- anthony_doan 7y ago> Is there a public link to your thesis? Not until my advisor okay my paper. He wants to see it first. > Can your approach be used to do feature segmentation of 3D (triangle mesh) data? Sorry, I have no clue what this mean so I don't think this is my area of expertise nor can I answer this properly. I'm an master student for applied statistic, this seems to be a computer science machine learning question? I can point toward papers regarding either GUIDE or CERP if you like.
- dewy 7y agoI'd love to read those papers if you wouldn't mind recommending them.
- anthony_doan 7y agoSelection Bias and GUIDE: http://pages.stat.wisc.edu/~loh/treeprogs/guide/wires.pdf http://pages.stat.wisc.edu/~loh/treeprogs/guide/wires.pdf GUIDE definitive paper: http://pages.stat.wisc.edu/~loh/treeprogs/guide/aoas260.pdf http://pages.stat.wisc.edu/~loh/treeprogs/guide/aoas260.pdf All other papers on GUIDE: http://pages.stat.wisc.edu/~loh/guide.html http://pages.stat.wisc.edu/~loh/guide.html CT CERP: http://www.ams.sunysb.edu/~hahn/psfile/aiim_moon.pdf http://www.ams.sunysb.edu/~hahn/psfile/aiim_moon.pdf