It is the era of computer-aided diagnosis, and ten thousand mammograms cross your system — but only about half a percent hide a genuine malignancy, and the costs of the two errors could not be more lopsided. Flag which scans warrant a radiologist's second read. A naive classifier maximizing accuracy will simply call everything benign and be 99.5% right while missing every cancer; meanwhile each false alarm spends a radiologist's scarce attention and, repeated, burns the trust that keeps them using the tool at all. Balance sensitivity against false alarms with the real asymmetric costs baked in. Get it wrong and you either miss the tumors the system exists to catch, or cry wolf so often the radiologists switch it off — extreme imbalance is the whole difficulty.
Chose The randomized tree ensemble (bagging & random forests) — wrong. The readable decision tree (CART) was the one that fit.
Breiman is the strongest fit on this roster, and on his own published record. CART (1984, with Friedman, Olshen, Stone) treats class priors and misclassification-cost matrices as core inputs — altering the 0.5% prior and penalizing a missed malignancy hundreds of times a false alarm is built into the splitting and pruning criteria, not retrofitted. Random forests (2001) added class-weighted and stratified-sampling variants aimed at exactly this skew, plus out-of-bag estimates for honest error assessment with few positives. His years as a freelance statistical consultant taught him messy, asymmetric real-world problems, and 'Statistical Modeling: The Two Cultures' (2001) is precisely an argument for judging models by predictive performance on the decision that matters rather than by fit. Feature engineering from images aside, this problem is nearly a Breiman worked example.
Rosenblatt's perceptron (1958) was the first trainable neural classifier, complete with a convergence theorem and hardware that learned visual patterns from examples — genuinely the right species of machine for this task, four decades early. But the perceptron's error-driven rule is exactly the algorithm the scenario warns about: with 0.5% positives, updates are dominated by the benign class, and a linearly inseparable, subtle radiographic boundary defeats it outright. Rosenblatt had no cost matrix, no ROC calibration, no way to weight a missed tumor two hundred times a false alarm, and 1950s-60s compute could not process real mammographic images. His probabilistic ambitions for the perceptron were sincere but mathematically thin, as Minsky and Papert demonstrated. He founded the lineage that eventually solved this problem while lacking nearly every specific instrument the solution required.
Battle #135 · 8/10/2026, 11:39:26 AM · this result is deterministic: the same two personas on this problem always resolve the same way.