AIFairnessModelBiasChecker¶

The AIFairnessModelBiasChecker class is a subclass of BaseModelBiasChecker in the Python package. This class is used for detecting bias in machine learning models and their predictions by using fairness metrics.

Configuration¶

Required Configuration¶

The AIFairness model bias checker requires the following configuration:

model_target: The model target, or label.
favorable_classes: values of the model_target that are considered favorable.
protected_attribute_names: The features that you want to check for bias.
privileged_classes: Values of the protected_attribute_names that you believe might be privileged.
privileged_groups: List of groups to consider privileged.
unprivileged_groups: List of groups to consider unprivileged.

Optional Configuration¶

The AIFairness model bias checker has no optional configuration.

Default Configuration¶

The AIFairness model bias checker has no default configuration.

Sample Configuration¶

  model_bias_checker: 
    config: 
        #the features that you want to check bias on
        protected_attribute_names : ["GENDER", "VACCINATED", "STERILIZED"] 
        # the values of the target variable that are considered good/favorable
        favorable_classes : [0,1]
        # the values of the protected_attribute_names that you believe may be privileged. 
        # Same order as above and each column can take a list of values
        # this should be something like [[1],[1],[1]] 
        privileged_classes:  [[1],[1],[1]]
        # list of groups to consider privilege/unprivilege. 
        # Each "group" is a dict of values of protected_attributes
        # 'AND' logic is applied between members of a dict, and 'OR' between dicts in the list
        #GENDER=1 AND VACCINATED=1 AND STERILIZED=1
        privileged_groups :   [{"GENDER" : 1, "VACCINATED":1, "STERILIZED":1}] 
        #GENDER=0 OR VACCINATED=0 OR STERILIZED=0
        unprivileged_groups : [{"GENDER" : 0}, {"VACCINATED":0}, {"STERILIZED":0}]

Methods¶

check_model_bias¶

Uses the provided data and model to check for bias in the model's predictions.

def check_model_bias(self, data, model, *args, **kwargs)

Arguments:

data (dict): A dictionary containing the training/testing data.
model (object): The model trainer object being checked for bias.

Returns:

metrics_out (dict): A dictionary containing the bias metrics computed by the function.

_eval_classes¶

This method evaluates the provided class list and returns it as a new list with evaluated elements. This function is called by the check_model_bias function when processing classes for fairness metrics.

def _eval_classes(self, class_list)

Helper function that evalutes lambda function defined in configuration into proper python functions.

Arguments:

class_list (list): A dictionary containing the training/testing data.

Returns:

class_list (list): A dictionary containing the training/testing data.

Usage¶

from lolpop.component import AIFairnessModelBiasChecker
import pandas as pd 

#get data
data_splitter_config = {
    #insert component config
} 
data_splitter = LocalDataSplitter(conf=data_splitter_config)
df = pd.read_csv("/path/to/data.csv")
data = data_spliter.split_data(df)

#get model
hpt_config = {
    #insert component config here 

}
oht = OptunaHyperparameterTuner(conf=hpt_config)
model = oht.run_experiment(data, model_version)

config = {
  #insert component config here
}
aifmbc = AIFairnessModelBiasChecker(conf=config)
bias_metrics = aifc.check_model_bias(data, model)