Building an ideal ML pipeline can be challenging in different ways. Maybe one of the most encountered challenges is data preparation. ML algorithms learn from the data you prepare for them. Without that data, those algorithms are practically useless. If we think of ML algorithms as little children, we can say that they simply imitate what they see in you and learn from you. You cannot expect any children to spot the wrongs in what they see for the first time. In this context, if the data you prepared for training the algorithm does not represent the current state of the world, the decisions it will make will not fit the goals set for the future. To be able to trust the decision-making ability of these algorithms, we need to make sure that they are trained in harmony with our goals.

As we all know, Machine Learning algorithms use past data to predict the future. But the past we need to look at may not be that far and if we know where to draw the line, it can turn ML into something more than a decision-maker into a gateway to a better future. To illustrate one of the pitfalls encountered while using historical data, we can concentrate on a popular issue: Credit Score.

Let’s say you have an idea of starting a business and don’t have an inheritance from a wealthy aunt or any investors, the only way to do so is to apply for a loan. The bank you applied to probably uses AI to help decide on the credibility of a person. The bank has put a ban on using features like race, gender, national origin and religion so that their algorithms are using the information about the person’s assets, total income, credit history, transaction analysis and work experience. They tell their algorithm that a person who has lots of assets, high income, clean credit history, unsuspicious transactions and nice-looking work experience can have a high credit score and if the case is all these are reversed, that person is supposed to have a low credit score. In a world full of rainbows and unicorns, this perspective can give justified answers but unfortunately, it is not the case.

The data that credit scoring models used are influenced by a generational wealth that many Black and Hispanic borrowers did not have equal access to, says Frederick Wherry, professor of sociology and director of the Dignity and Debt Network at Princeton University. Statistics show that only 1 in 5 Black consumers and 1 in 9 Hispanic consumers are above the average credit scores, while 18 out of 19 white consumers are above the average credit scores. This may lead us to think whether the algorithms that make these decisions about credit scores are unnecessarily influenced by the past. So at this point, the question should be: “How to prevent ML from being stuck in the past?”

Unfortunately, not using the features like race or religion does not help the decision become unbiased. In machine learning, the features may have a stronger dependency than expected. For any other feature than race, removing the one with less relevance solves the problem. In this case of credit scoring, remaining features like income still carry the racial bias behind and it is unwanted. Rather than looking back at the past, it may give a better idea to check the current situation and then predict the future.

There may be more than one way to fix this issue for good but the most realistic and applicable one is suggested as using cash flow and payment history on rent and utilities as features. Cash-flow underwriting is based on how much money is in your bank account each day over the last year so it is a feature that helps check the near past of consumers.

One may think that companies have to compromise about their profits to become unbiased but looking at the empirical research by FinReg Labs proves the opposite. FinReg Labs found compelling evidence that shows among the sample populations, cash-flow variables and scores are predictive of credit risk across the highly heterogeneous set of research participants. According to the study, the cash-flow metrics were both predictive on their own and also frequently improved the ability to predict the credit risk in combination with traditional credit scores. This study invalidates the preconceived notion that unbiased credit scoring models require compromises from business goals. The solution seems to be more about keeping the algorithms up to date.

All in all, in a constantly changing world, the models that we trust with our decisions should be adaptive and present. To make sure that ML algorithms are using the right data that represents the current situation of the problem, we need to increase their capabilities of them. If we use the features that are constantly updated like cash-flow underwriting with traditional machine learning, as the new data emerges, data scientists retrain the models and most probably change the hyperparameter settings to get good results with the new dataset. At this point, Adaptive Machine Learning holds the key to a better future. This advanced concept enables the models to follow up on every feedback provided to make future predictions even better. With this agility, it becomes easier for businesses to receive more useful decisions from machines. If only the ones that adapt can survive, then our ML should also be able to adapt to this changing world.


[1] Yapo, Adrienne, and Joseph Weiss. “Ethical implications of bias in machine learning.” Proceedings of the 51st Hawaii International Conference on System Sciences. 2018.

[2] Jiang, Heinrich, and Ofir Nachum. “Identifying and correcting label bias in machine learning.” International Conference on Artificial Intelligence and Statistics. PMLR, 2020.