All MCQs

Nearby or strongly connected observations are assumed likely to share labels.
The graph structure helps exploit unlabeled data.

Methods such as Platt scaling map decision scores to probabilities.
Calibration should use separate or cross-validated data.

Each classifier separates one target class from the rest.
The class with the strongest decision score is often selected.

Multiple binary classifiers are trained and their outputs combined.
Libraries typically implement one of these strategies automatically.

Large-scale features can dominate the geometry of the optimization.
Consistent scaling also makes C and gamma tuning more meaningful.

Very localized kernels can memorize training details.
Validation is needed to balance flexibility and generalization.

Large gamma gives narrow local influence and complex boundaries.
Small gamma creates smoother, broader effects.

The radial basis kernel measures localized similarity between points.
Combining support-vector influences yields curved decision regions.

The linear kernel uses the ordinary dot product.
It is effective for many high-dimensional sparse problems.

A kernel enables nonlinear boundaries without explicitly constructing all transformed features.
The SVM optimization uses pairwise kernel values.

A large penalty forces the model to avoid training violations aggressively.
This can make the boundary sensitive to noise.

Large C emphasizes fitting training data, while small C permits more violations for a wider margin.
It controls regularization strength inversely.