README.md

Bonsai, a library to parse and evaluate decision trees encoded as JSON

This code allows to export Gradient Boosted Tree Model generated by gbm into a file in JSON format and to evaluate them on new data outside R.

We have been using it inside our company to deploy gbm-trained model into production, either evaluating the JSON directly using the scala code or by generating C++ code from the JSON and deploying the generated code.

Instructions

You can follow the script under src/test/R/test.R' to see how to create a gbm model and convert it to JSON using the script under src/main/R/gmb.R`. The example uses the the model from gbm's help page (except I reduced the number of trees from 1000 to 100 in order to keep the files small) to prepare a test case for the scala unit testing code.

Notes

gbm handles ordered categorical factors as numeric variables and it requires careful handling when converting to JSON. I actually represent the conditional as set operators in the JSON files in order to preserve the categorical nature of the variables.

Dependencies:

• play-json

Steps:

• Build the package: mvn clean package

• Run the Bonsai tool: java -jar target/bonsai-1.0-jar-with-dependencies.jar

  Usage: to pretty-print a json moodel: Bonsai -p modelFile to evaluate a model over a test file: Bonsai -e modelFile dataFile to generate C++ code: Bonsai -c (modelFile)*

• To convert the model json file to cpp file:

  java -jar target/bonsai-1.0-jar-with-dependencies.jar 'model1 json file' 'model2 json file' ...

Contributors

• Alex Cozzi
• Mirek Melichar
• Amit Jaiswal
• mijiang
• Marco Pennacchiotti