dongdyang/shapley

Shaple Value Regression with Gini Coefficients

Using Shapley Value Regression with Gini coefficients

This is code for running Shapley Value Regression with Gini coefficients. After training a model, we want to learn how each predictor contributes to the model prediction. One way to tackle this problem is to use Shapley Value Regression (SVR). A conventional SVR approach applies on a linear regression model and decompose the R-square using Shapley values. This code uses Gini coefficients instead of R-square so that it can be applied to not only linear regression models but also a variety of different models such as decision tree, neural network, etc.

Current Supported Models

Currently, this code only supports a linear regression model and a log-linear regression model. The log-linear regression model is the model in which the dependent variable is transformed using the logarithm function. In the future, more models will be added.

Usage

usage:

train.py [-h] [-v] [-i INPUT] [-o OUTPUT] [-m MODEL] [-s STEP]

optional arguments:
  -h, --help            show this help message and exit
  -v, --verbose         verbose computation
  -i INPUT, --input INPUT
                        add the input csv file
  -o OUTPUT, --output OUTPUT
                        the csv file you want to output (default result.csv)
  -m MODEL, --model MODEL
                        specify the model, currently support 'linear' or 'log-
                        linear', default: linear
  -s STEP, --step STEP  steps for computing each Shapley value (default
                        10,000)

Training

1. Prepare your training data and record the path of your data

You can create a directory named "data" in the root directly. I created a "test_data" directory for you to try and test this code.

The data should be saved in a csv file. The first row should be the name of each variable and the first column should be your dependent variable.

2. run the "train.py" using python3.

Please specify the input and output files using "-i" for input and "-o" for output. The input path is the path of your csv file. The default output is "result.csv" in the root directory.

3. specify the model you use by "-m". Currently, only "log-linear" and "linear" are supported. The default value is "linear".

Using The Code of the Shapley Regression Directly

You can also using Shapley regression directly without running in the command line. You can import the shapley model in your code:

import shapley  # make sure your file is at the root directory

Requirements

Numpy

Sklearn

TODO

• Neural Network Model