Overview
Classification model development fails in two predictable places: metric selection and threshold calibration. Accuracy is the default metric — and the wrong metric for any imbalanced problem. A model that predicts "no fraud" for every transaction achieves 99.8% accuracy on a dataset where 0.2% of transactions are fraudulent. The threshold of 0.5 is the default — and the wrong threshold for any problem where false positives and false negatives have different costs.
The right classification framework starts with the cost structure: what is the relative cost of a false positive vs. a false negative? The answer determines the evaluation metric, the decision threshold, and the model selection criterion. Everything else follows.
The Classification Model Framework Prompt generates a complete classification specification: cost-structure-driven metric selection, algorithm selection by data properties, class imbalance handling, threshold calibration, and a business-impact evaluation that translates model performance into the units that matter to the decision-maker.
What you get: - Cost-structure-driven metric selection - Algorithm selection matrix by data properties - Class imbalance handling strategy - Decision threshold calibration - Business-impact performance evaluation
Built for: data scientists and ML engineers building classification models where the cost of errors is asymmetric and accuracy is not a sufficient metric.