Source code for the paper: Vanishing Boosted Weights (VBW): A Consistent Algorithm to Learn Interpretable Rules

This Python package allows to reproduce the results presented in the paper "Vanishing Boosted Weights: a Consistent Algorithm to Learn Interpretable Rules". The potential users can run the new method on their own data.

The user can see the user manual by executing python3 manager.py -h.

usage: manager.py [-h] [-a ALGORITHM [ALGORITHM ...]] [-f FEATURES]
                  [-e ESTIMATORS [ESTIMATORS ...]] [-d DATA] [-p PROCESS]

Vanishing Boosted Weights (VBW): A corrective fine-tuning procedure on
decision stumps.

optional arguments:
  -h, --help            show this help message and exit
  -a ALGORITHM [ALGORITHM ...], --algorithm ALGORITHM [ALGORITHM ...]
                        List of arguments (default: GBoost CatB GOSS VBW
                        LightGBM Averaged])
  -f FEATURES, --features FEATURES
                        Number of features (default: 10)
  -e ESTIMATORS [ESTIMATORS ...], --estimators ESTIMATORS [ESTIMATORS ...]
                        List of number of estimators (default: 1 5 10 25 50 75
                        100)
  -d DATA, --data DATA  Path to datasets (default: ./Examples)
  -p PROCESS, --process PROCESS
                        Number of processes (default: 4)

The list of learning approaches are:
GBoost : Gradient Boosting Decition Trees
CatB : CatBoost Classifier
LightGBM : Light Gradient Boosting Machine
GOSS : Gradient-Based One Side Sampling
VBW : Vanishing Boosted Weights
Averaged : Corrective Federated Averaging VBW

Below is an example of using the package:

python3 manager.py -a VBW GBoost GOSS Averaged -f 10 -e 1 10 25 75 90 -d Examples -p 15 

A dataset is a text file comprising of all the samples. Each sample is represented by a line of features separated by space. The last column is the class which is either 1 or -1.