Logistic Regression

On the Wisconsin breast-cancer measurements that ship with scikit-learn, 569 tumours with 30 numbers each, LogisticRegression() estimates the probability that a tumour is malignant with 97.9% five-fold cross-validated accuracy and a log loss of 0.074, the features standardised inside a pipeline. That default includes a penalty nobody asked for. Switch it off with penalty=None and the solver stops at its 100-iteration limit with a ConvergenceWarning. Allow it 100,000 iterations and it reports success, with every training tumour classified correctly and a weight vector 27,225 units long. Cross-validate an almost unpenalised model (C=10000) and the held-out log loss is 1.03, fourteen times the default's, although accuracy has fallen only three points.