ig-cd/decision-stream

Clojure implementation of the paper "Decision Stream: Cultivating Deep Decision Trees"

Decision Stream

1. Overview [paper]

This repository provides a basic implementation of the Decision Stream regression and classification algorithm. Unlike the classical decision tree approach, this method builds a directed acyclic graph with high degree of connectivity by merging statistically indistinguishable nodes at each iteration.

2. Prerequisites and Dependencies

• Clojure
• Apache Commons Math
• JBLAS (requires ATLAS or BLAS/LAPACK)
• OpenCSV

The dependencies are configured in the pom.xml file.

3. First steps

• Extract the archive data.gz with training data by running tar -xvzf data.gz
• Optional: rebuild decision-stream.jar with Leiningen (lein uberjar) or Maven (mvn package).

4. Train the model

java -jar decision-stream.jar base-directory train-data train-answers test-data test-answers learning_mode significance-threshold

The program takes 7 input parameters:

base-directory   -   path to the dataset
train-data   -   file with training data
train-answers   -   file with training answers
test-data   -   file with test data
test-answers   -   file with test answers
learning_mode:   classification or regression   -   classification or regression problem
significance-threshold   -   threshold for merging/splitting operations

Example:

java -jar decision-stream.jar data/ailerons/ train_data.csv train_answ.csv test_data.csv test_answ.csv regression 0.02

5. Provided datasets

The datasets prepared for training in the data folder:

• Ailerons (F16 aircraft control problem)
• Credit scoring
• MNIST