Logistic Regression Regularization — Reduce Overfitting

Quick Answer

A training accuracy of 0.99 against a validation accuracy of 0.75 is the textbook signature of overfitting: the model has essentially memorized patterns specific to the training set, including noise, rather than learning the general relationships that hold on unseen data, and with 100 features relative to the dataset, it has plenty of room to do so. Increasing the regularization strength addresses this directly by adding a penalty on the size of the model's coefficients, which discourages the model from assigning large weights to features that only help it fit training-set quirks, effectively simplifying the decision boundary so it generalizes better. This is a more targeted fix than something like collecting more data or changing the model architecture, because the failure mode described, a huge accuracy gap between training and validation with no other symptoms mentioned, points specifically at model complexity rather than data quantity or algorithm choice, and regularization is the standard lever built into logistic regression for controlling that complexity. Whenever a question shows a large gap between training and validation performance with everything else about the setup held constant, that gap is the signal to look for an answer that constrains the model rather than one that adds capacity or more training time.