Loan Approval
Bias Mitigation for AI/ML Applications

This project examined the impact of using biased data that could be used to train algorithms associated with learning from credit-based data sets.

Industry-defined fairness metrics were used to assess whether bias was present in the data. A bias-mitigating technique was then applied to produce a dataset without bias.

Role: Data Scientist

The Outcomes

Enhanced Fairness in
Loan Approvals

The implementation of bias mitigation techniques resulted in a fairer loan approval process. By addressing biases related to age and gender, the project ensured more equitable treatment of all applicants, leading to increased trust and fairness in the decision-making process.

Improved Accuracy of
AI Models

Through the application of bias detection and mitigation, the accuracy of AI models was significantly enhanced. This improvement in model accuracy helped in making more reliable predictions, ultimately reducing the risk of incorrect loan approvals and denials.

Strengthened Ethical
Compliance

The project reinforced ethical standards by integrating fairness metrics into the data processing pipeline. Ensuring that the AI models adhered to these ethical guidelines minimized discrimination, promoting ethical AI usage and compliance with regulatory standards.

The Tools

The Process

Banks use AI and Machine Learning to predict whether borrowers will repay their loans, but if the data is biased, it can lead to unfair decisions. This project addresses the challenge of mitigating bias present in loan approval datasets. The goal is to identify and reduce the bias, ensuring that AI-driven credit decisions are fair and trustworthy.

Raw Data Collection and Preparation

The project began with gathering raw data and performing an initial data cleansing to make the data easier to work with. This includes standardizing the data, converting columns to the correct types, and removing unnecessary information. Clean data was crucial for accurate analysis and reliable results.

To address several missing values in the dataset, a method called KNNImputer is used for numerical data. This method estimates missing values based on similar entries, providing a more realistic and accurate way to handle gaps in the data.

Data Cleansing

Data Preprocessing

Industry-standard fairness metrics were used to assess the dataset for bias. Specifically, metrics such as Statistical Parity Difference (SPD) and Disparate Impact (DI) were used to evaluate potential biases in relation to protected class attributes like gender and age. This step was essential to highlight potential biases in the data, ensuring that any discriminatory patterns were flagged for further processing. Upon detecting bias, a pre-processing algorithm was applied to mitigate discrimination across gender and age before splitting the data into training and testing sets.

Bias Identification

Training and Testing

The training data underwent standard training and in-processing algorithms to enhance fairness during model training. The testing data was used to evaluate the model's performance, focusing on both accuracy and fairness metrics. This phase ensured that the AI models were trained to be both effective and equitable.

Bias Mitigation

Post-processing Bias Mitigation and Model Evaluation

Post-processing algorithms were applied to further reduce any remaining bias after model training. The Disparate Impact Remover (DIR) algorithm from the AI Fairness 360 (AIF360) toolkit was utilized to adjust the distributions of privileged and unprivileged groups for each feature to be more similar. The classifier's performance was assessed for both accuracy and fairness using classifier unit tests and dataset metrics.

Deployment and Monitoring

After ensuring the model met the required fairness and accuracy standards, it was ready for deployment. The model's performance and fairness were continuously monitored, with external interventions applied as necessary to reprocess and retrain the model to maintain fairness. This ongoing monitoring was crucial for ensuring the responsible use of AI in loan approvals.