https://github.com/astupidbear/mlinterpret.jl

A Meta Package for Machine Learning Interpretation
https://github.com/astupidbear/mlinterpret.jl

Last synced: 7 months ago
JSON representation

A Meta Package for Machine Learning Interpretation

• Host: GitHub
• URL: https://github.com/astupidbear/mlinterpret.jl
• Owner: AStupidBear
• License: mit
• Created: 2019-10-25T16:40:52.000Z (almost 6 years ago)
• Default Branch: master
• Last Pushed: 2021-09-08T07:15:24.000Z (about 4 years ago)
• Last Synced: 2025-01-20T20:27:07.077Z (9 months ago)
• Language: Julia
• Homepage:
• Size: 546 KB
• Stars: 22
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# Machine Learning Interpretation

[![Build Status](https://github.com/AStupidBear/MLInterpret.jl/workflows/CI/badge.svg)](https://github.com/AStupidBear/MLInterpret.jl/actions)

[![Coverage](https://codecov.io/gh/AStupidBear/MLInterpret.jl/branch/master/graph/badge.svg)](https://codecov.io/gh/AStupidBear/MLInterpret.jl)

## Installation

```julia

using Pkg

pkg"add MLInterpret"

```

Try without installation using docker

```bash

docker run -it --rm astupidbear/mli

```

Or build it from Dockerfile

```bash

url=https://raw.githubusercontent.com/AStupidBear/MLInterpret.jl/master/Dockerfile.py

python3 -c "$(curl $url)"

```

## Usage

```julia

using MLInterpret

using PyCall

using PyCallUtils

using PandasLite

X = DataFrame(randn(Float32, 10000, 5))

y = (X[3] > 0) & (X[2] >= 0)

@from lightgbm imports LGBMRegressor

model = LGBMRegressor()

model.fit(X, y)

```

You can interpret any machine learning model from Python which has a property `.predict` by calling

```julia

interpret(model, X, y)

```

If your model dosen't have a property '.predict' (like Julia models), you can still interpret its predictions by

```julia

ŷ = model.predict(X)

interpret(X, ŷ)

```

This will generate a folder `mli` in the current directory which contains

- `pdp.pdf`: partial dependency plot [link](https://oracle.github.io/Skater/reference/interpretation.html#partial-dependence) ![](screenshots/pdp.png)

- `perturb_feaimpt.csv`: feature importance calculated by purturbation [link](https://oracle.github.io/Skater/reference/interpretation.html#feature-importance) ![](screenshots/perturb_featimpt.png)

- `shap.pdf`: shap value [link](https://github.com/slundberg/shap) ![](screenshots/shap.png)

- `shap2.pdf`: shap interaction value [link](https://github.com/slundberg/shap) ![](screenshots/shap2.png)

- `surrogate_tree-*.pdf`: surrogate tree [link](https://oracle.github.io/Skater/reference/interpretation.html#skater.core.global_interpretation.tree_surrogate.TreeSurrogate) 

![](screenshots/surrogate_tree.png)

- `actual.pdf`: actual plot [link](https://pdpbox.readthedocs.io/en/latest/actual_plot.html) ![](screenshots/actual.png)

- `actual2.pdf`: actual interaction plot [link](https://pdpbox.readthedocs.io/en/latest/actual_plot_interact.html) ![](screenshots/actual2.png)

### MLI with [H2O Driverless AI](https://www.h2o.ai/products/h2o-driverless-ai/)

#### Start DAI

```bash

docker run -d \

    --pid=host \

    --init \

    -u `id -u`:`id -g` \

    -p 12345:12345 \

    -v /dev/shm:/dev/shm \

    astupidbear/dai:1.7.0

```

You can get a trial license of H2O Driverless AI from [H2O](https://www.h2o.ai/try-driverless-ai/), then open `http://127.0.0.1:12345/`, login and enter your license.

#### Interpret

```julia

dai_interpret(X, y)

```

Open `http://127.0.0.1:12345/`, click `MLI`, choose the toppest `Interpreted Model`

![](screenshots/dai.png)

### MLI with [Bayesian Rule List](https://oracle.github.io/Skater/reference/interpretation.html#skater.core.global_interpretation.interpretable_models.bigdatabrlc.BigDataBRLC)

#### Installation

```julia

using MLInterpret

MLInterpret.install_brl()

```

### Interpret

```julia

sbrl_interpret(X, y)

```

A file named `sbrl.txt` will be created in your working directory.

![](screenshots/sbrl.png)