Page 9 - summer of ai

SUMMER OF AI JULY 2019 DATA & ANALYTICS
NeuroDecision® plus differentiated data generated
25%h
Uplift in KS
for one of the largest
US banks, resulting in
the potential approval of 500,000 new clients every year while keeping risk levels constant
input signals into new signals that are further passed to other neurons. Each neuron can be thought of a logistic regression model. Instead of relying on guesswork to divide the modeling work among multiple models (or neurons),
we simultaneously optimize all input weights to every neuron to minimize overall error on training data.”
In other words, data scientists employ neural networks to combine input signals (also known as attributes) into signals that trigger neurons, which also emit signals that are then forwarded and combined to make a credit decision. The decision could be “What's the probability that you pay your account as agreed in the terms over a two-year period or not?”
Unlike traditional logistic regression models, neural network models are “universal approximators.” The modeling technique has a feature that allows data scientists to extract non-linear trends and attribute interactions from the data automatically. A neural network automatically fits the non-linear data and finds its interactions. In contrast, data scientists must work very hard to explicitly incorporate
non-linear trends or attribute interactions into a regression model.
However, the tradeoff with universal approximators is the inability to explain the key factors impacting the model. Neural networks are an extremely efficient and highly effective means of making decisions, and will approve more clients while maintaining a customer’s risk profile. The challenge with a neural network is understanding how these signals are combined to make a credit decision. From white papers to technology magazines, this challenge is commonly called “the black box.”
One of the ways Matt and his team of data scientists overcome the inability to explain is through the use of logical constraints to the modeling process.
What’s wrong with being perfect?
Contrary to popular belief, constraints often improve a model’s performance. In numerous trials over a three-year study, Matt observed the imposition of logical constraints actually improved model performance over time. One
well-documented challenge with any machine learning model is that it can easily overfit development data. Data scientists must be careful and impose some type of constraints on models so that the models do not become too perfect and unable to perform “in the wild.”
One such logical constraint is monotonicity.
In essence, monotonicity means the NeuroDecision algorithm always rewards positive behavior (score increases) and penalizes negative behavior (score decreases). When a borrower re-pays debt every month on time, NeuroDecision Technology identifies the behavior as positive, increases the score, and results in positive outcome: in many cases an offer of credit.
Superior explanations come from regulatory insights
An ongoing conversation between data scientists and regulators continues because regulators must understand how the AI techniques used at Equifax fit into existing regulatory requirements. The Data Science Lab socialized the concept of explainable AI with quite a few different regulatory agencies in the United States and abroad. The data scientists met many times with the Federal Reserve Board, Federal Communications Commission, the Federal Deposit Insurance Corporation,
the National Credit Union Association and the Consumer Financial Protection Bureau as well.
Aggregate explanations ignore the individual person
Competitors may boast their explainable AI techniques provide compliant explanations akin to the reason codes generated by NeuroDecision Technology. However, Matt argues only NeuroDecision Technology can provide logical and actionable explanations
9