Machine Learning Failure Modes: Why Production Models Break

This post details the diagnosis and mitigation of underfitting, including identifying it through high training and validation errors, understanding its relation to the bias-variance tradeoff, and outlining a systematic approach to fixing it by increasing model capacity, reducing regularization, engineering better features, improving data quality, and systematically tuning hyperparameters. It also highlights the importance of data leakage and shortcut learning as common failure modes in ML systems, emphasizing that production ML failures cluster into four categories: data issues, shortcut learning, system complexity, and organizational failures. The post advocates for proactive monitoring of input data distributions using statistical tests like Kolmogorov-Smirnov, Population Stability Index, and Jensen-Shannon Divergence to catch model degradation early, and stresses the importance of temporal validation and auditing features to prevent data leakage. It also introduces worst-group-accuracy as a metric to test for shortcut learning and discusses advanced techniques like counterfactual data augmentation and invariant risk minimization.