by Ruth Lee, CMB
There is so much to unpack with artificial intelligence, but one of the starting points is the responsible and ethical use of AI in mortgage banking, especially when it impacts any form of decision-making. Already states such as California, Colorado, and Connecticut are enacting laws that, while short on detail, have significant impacts on the use of AI in mortgage banking. Here in Colorado, the most recent law was signed with a lot of work still needed to ensure it doesn't have unintended consequences or too broad an interpretation, potentially impacting existing automation like Automated Underwriting Systems (AUS) and credit scoring.
Understanding Data Bias
Data is the lifeblood of AI. It's the raw material that powers what is, after all, artificial intelligence—with an emphasis on artificial. The quality of the output directly depends on the quality of the input data. Unfortunately, both historical and current data can bear the marks of biases. Regulatory frameworks like the Fair Housing Act, Fair Credit Reporting Act, RESPA, ECOA, and HMDA provide ethical guidelines, but our data still sometimes reflects these biases. It's akin to a historical record, where some chapters echo discriminatory practices of the past, and even present data can exhibit similar flaws. Thankfully, AI comes equipped with tools that can help us identify and address these biases, allowing us to pave the way for fairer and more equitable decision-making.
Before we can mitigate bias, we need to understand what it is. Data bias occurs when the data used to train AI models reflects existing prejudices or imbalances in society. This can happen due to a variety of reasons:
Historical Inequities: Past discriminatory practices can be embedded in historical data.
Sampling Bias: When the data collected is not representative of the broader population.
Measurement Bias: When the data collection process itself introduces errors or biases.
These biases can lead to AI models that perpetuate or even amplify existing inequalities.
Mitigating The Echoes of the Past
What is deeply exciting about AI's potential is that it offers a unique opportunity to fine-tune vast amounts of data to address these historical biases. This involves scrutinizing our data, identifying past discrepancies, and ensuring our AI systems are nurtured on unbiased information. So, how is that done... in reality?
Mortgage banking aims to connect investors with borrowers efficiently. Mitigating bias in AI training sets for mortgage banking is both an art and a science, demanding a blend of meticulous planning and advanced technical strategies. Let’s start with data collection and preprocessing, the first step in building an unbiased AI system. Imagine you're trying to create a fair lending process—this starts with collecting data that represents a wide range of people from various backgrounds and financial situations. This ensures the AI doesn't favor any one group over another. Standardizing this data—meaning putting it all into a common format—prevents inconsistencies that could lead to unfair outcomes. Additionally, anonymizing data by removing personal identifiers, such as names or social security numbers, helps ensure the AI is blind to factors like race, gender, or age, which shouldn't influence lending decisions.
Bias Detection and Data Augmentation
The next step is data augmentation and balancing. This involves creating a dataset where all groups are fairly represented, even those who might be underrepresented in the original data. One way to do this is by generating synthetic data. Think of this as creating artificial but realistic data points that fill gaps in your dataset. Techniques like SMOTE (Synthetic Minority Over-sampling Technique) can generate new data points that reflect the characteristics of underrepresented groups. Additionally, balancing techniques like oversampling, which involves making more copies of the data from underrepresented groups, and undersampling, which involves using fewer copies of the data from overrepresented groups, help ensure the dataset is balanced. This is like making sure a choir has enough tenors and basses, not just sopranos and altos, so the harmony is balanced.
Explainable AI (XAI) and Ongoing Monitoring
When it comes to designing and training algorithms, fairness-aware techniques are essential. Algorithms need to be constructed in a way that inherently promotes fairness. Adversarial debiasing is one method, where the model is trained to minimize bias actively. Fair representation learning is another technique that transforms features in a way that removes bias. Regularization techniques help prevent overfitting—this is when the model learns too much from the training data and doesn’t perform well on new data. It's like preparing a student not just for one specific test, but for a wide range of questions they might encounter in the future. Bias mitigation techniques during training, such as reweighting data points to ensure all voices are equally heard or modifying the loss function (which measures the model’s performance) to penalize biased outcomes, further enhancing fairness.
Evaluating and validating the model’s performance is another critical step. This involves using diverse test sets that represent different populations to ensure the model performs fairly across all groups. For example, you wouldn’t test a new medical treatment only on young, healthy individuals—you’d want to see how it works across all ages and health conditions. Incorporating fairness metrics alongside traditional performance metrics provides a comprehensive view of the model’s fairness. Techniques like cross-validation, where the data is divided into subsets and the model is tested on each subset, help ensure that the model is consistent and fair. Continuous monitoring and periodic audits of the model ensure that it remains unbiased over time. Establishing feedback loops allows users and stakeholders to report biased outcomes, which can then be addressed by retraining the model and updating the data.
By integrating these practices, businesses can better navigate the complexities of data bias and ensure their models perform fairly and accurately across all user demographics. As David Coleman, President of MISMO, notes, “MISMO is bringing together industry professionals to share knowledge, collaborate on potential industry use cases, and develop standards in this rapidly developing field. The initial focus will be on creating a shared vocabulary through a glossary of AI terms that are applicable to the mortgage industry.”
Finding Ethical Use in Mortgage Banking AI
In conclusion, mitigating bias in AI training sets for mortgage banking involves a comprehensive approach that spans from initial data collection to ongoing monitoring and adjustment. By ensuring data diversity, balancing datasets, designing fairness-aware algorithms, and thoroughly evaluating and validating model performance, we can create AI systems that make fair and unbiased decisions. This not only fosters trust and reliability in AI-driven mortgage processes but also supports a more equitable financial landscape for everyone. As Rohit Chopra, Director of the CFPB, emphasizes, “Companies must take responsibility for the use of these tools. Unchecked AI poses threats to fairness and our civil rights.”
Resources:
IEEE Innovation at Work, "Understanding the Ethics of Artificial Intelligence" (IEEE Innovation at Work)
IEEE Innovation at Work, "How Can Organizations Ensure Their AI Systems Are Ethical?" (IEEE Innovation at Work)
IEEE Innovation at Work, "Improving Ethics in AI" (IEEE Innovation at Work)
ZDNet, "Consumers More Skeptical on AI Ethics"
Deloitte Insights, "Ethics and the Future of AI"
Forbes, "How Do We Use Artificial Intelligence Ethically?"
Comments