A support vector machine trained to predict diabetes risk from clinical measurements — the kind of small, tabular medical dataset where getting the fundamentals right matters more than reaching for a bigger model.
Data and features
768 patient records, each with clinical measurements like glucose level, BMI, blood pressure, and age. Every feature ran through StandardScaler normalization before training — SVMs are distance-based under the hood, so leaving features on wildly different scales would have let high-magnitude ones quietly dominate the decision boundary.
Model selection and tuning
With a dataset this size, a support vector classifier was a better fit than anything deep-learning-shaped — enough structure to find a real margin, not enough data to justify the variance a bigger model would add. Stratified train/test splits kept the class balance honest, and grid search swept the kernel, C, and gamma hyperparameters to find the combination that generalized best rather than the one that just fit training data hardest.
Reading the results
The gap between the two is narrow, and that gap is the signal that matters more than either number alone. A model that scores much higher on training data than on test data is memorizing, not learning; this one wasn't.