mini-pac-bayes-tree

Essential files to construct the pac-bayes-tree and compare its test-set performance to that of some common tree-pruning approaches. Please refer to https://papers.nips.cc/paper/8158-pac-bayes-tree-weighted-subtrees-with-guarantees.pdf for details on definitions of penalty and priors.

Example uses

pac-bayes-tree:

Example call and meaning of arguments

python dataMaster_[ADDT_nS_master]_[eTS,MD].py 0 0 leakyAda effInt penFirst spam

1st argument: index into array of training sizes 2nd argument: index into array of training/testing splits 3rd argument: penalty type 4th argument: prior type 5th argument: order of optimization to tune hyper-parameters
6th argument: dataset to experiment on

Where are results saved?

For instance, "eTS/each experiment, each train size/spam-ADDT_nS_leakyAda_effInt-ex0tr0.npy" will store the test error of the pac-bayes-tree.

tree-pruning

Example call and meaning of arguments

python dataMaster_[PkdDT_master]_[eTS,MD].py 0 0 effInt spam

1st argument: index into array of training sizes 2nd argument: index into array of training/testing splits 3rd argument: prior type 4th argument: dataset to experiment on

Where are results saved?

For instance, "eTS/each experiment, each train size/spam-PDDT_effInt-ex0tr0.npy" will store the test error of the tree-pruning.

Remarks

Computational efficiency

pac-bayes-tree approaches have the same complexity as tree-pruning in the sense that the number of internal nodes visited at training/test time are roughly the same. The current implementation of pac-bayes-tree is slower than tree-pruning because of the extra precision necessary to compute weights for pac-bayes-tree.