How a Neural Network Learned Its Own Fraud Rules: A Neuro-Symbolic AI Experiment

A neuro-symbolic AI experiment demonstrates how a neural network can learn its own auditable IF-THEN fraud rules directly from data. The proposed architecture uses a hybrid model with a standard MLP for detection and a parallel path for differentiable rule induction, which learns to explain the network's predictions. This rule-learning path consists of a learnable discretizer and a rule learner layer that transform continuous features into symbolic logic. When tested on a credit card fraud dataset, the model successfully learned interpretable rules and independently rediscovered a key feature known to be highly correlated with fraud.