Bias Mitigation#

Objectives: what you will take away#

Definitions & Understanding of data bias and why it occurs.
How-To perform bias mitigation using Howso Engine.

Prerequisites: before you begin#

Installation

You have successfully installed Howso Engine
You have loaded, configured, trained, and analyzed data
You are familiar with feature contributions

Notebook Recipe#

The following recipe will supplement the content this guide will cover:

Bias Mitigation

Concepts & Terminology#

What is biased data?

Often, poor data quality and subsequent unequitable analysis results arise from biases in data. There are a variety of ways bias may affect your data:

Sampling and representation bias - lack of diversity of the populations contained in the data
Measurement and omitted variables bias - choice of information contained in the dataset
Aggregation bias - data is generalized across populations
Linking bias - various datasets are combined

Biases in data may propagate through data analytics and AI/ML models, leading to biased results upon which predictions and decisions might be made. Additionally, biased results may be projected into the future and lead to missed opportunities to create true value. This potentially harms various groups of people and reduces the overall utility of the data. Because analytics and modeling results that use biased data are increasingly falling under regulations, decisions that are based on insights gained from biased data might lead to breaches of these regulations.

In addition to understanding bias, we recommend being familiar with the following concepts:

How-To Guide#

There are two ways to use Engine to understand and mitigate data bias.

Method 1 - Explore data for bias#

Detect which protected attributes in a dataset contribute the most to model predictions, showing biases users may or may not have recognized.

After building, training, and analyzing a Trainee, you can use the details argument of react to obtain the mean decrease in accuracy of including a feature and feature contributions to understand which features are driving a prediction.

# Details describe the information we are getting from a given react call
details = {
    'feature_mda': True,
    'feature_contributions': True
}

# Obtain indices and session ids
stored_data = t.get_cases(
    session=t.active_session,
    features=df.columns.tolist() + ['.session_training_index', '.session']
)

# React
results = t.react(test_case[context_features],
                context_features=context_features,
                action_features=action_features,
                details=details
)

Step 2 - Control predictions for data bias#

Easily update which features are used in predictions, so features that contribute to bias can be omitted. Within the react call, specify only the context features that do not contribute to prediction bias.

# React
results = t.react(test_case[context_features], # Input features - only those which do not include bias
                context_features=context_features, # Input features - only those which do not include bias
                action_features=action_features,
                details=details
)