kenkyusha / sdt

Soft Decision Tree

This is reimplementation of Ozan Irzoy Soft Decision Trees( http://www.cs.cornell.edu/~oirsoy/softtree.html ) in Matlab.

exampleData.mat is toy dataset which can be used to train the tree.

commands:

b = SoftTree(trainData,trainTarget,trainData,trainTarget); b.train()

Following parameters are scattered around Node.m and set to these default values.

HARDINIT true // if true, optimization starts from hard tree parameters, else, randomly

MINALPHA 1 // starting range of learning rate

MAXALPHA 10 // ending range of learning rate

MAXEPOCH 25 // number of epochs in training

MAXRETRY 10 // number of restart of optimization from a starting point

PRETH 1e-3 // prepruning threshold

TODO:

• port the code into python
  • http://mathesaurus.sourceforge.net/matlab-numpy.html
• example comparison
  • decision tree
  • random forest
  • soft decision tree
• compare
  • depth
  • complexity