A supervised learning method that finds a decision boundary (hyperplane) separating classes by maximizing the margin between support vectors—the training points nearest the boundary—optionally mapping inputs into a higher-dimensional feature space via kernel functions to handle nonlinearity.

Principle

Formulate classification as a convex optimization: maximize the margin subject to classification constraints (soft margin permits slack for misclassification with regularization), or use kernelized representations to perform the margin maximization in an implicit feature space without explicit coordinates.

Demonstration

Binary classification on linearly separable data: the SVM finds the hyperplane equidistant from the nearest examples of each class; with nonseparable data, a soft‑margin SVM allows misclassifications while penalizing them via a regularization parameter; with an RBF kernel, nonlinear class boundaries can be learned implicitly.

Misapplication

Using SVMs without feature scaling, choosing inappropriate kernels or hyperparameters blindly, applying SVMs to extremely large datasets without approximation, or interpreting the margin as calibrated probabilities without post‑processing (SVM outputs are not probabilistic by default).

Consequence

Yields robust classifiers with good generalization when margin and kernel are well-chosen; supports sparse solutions determined by support vectors, but performance depends on kernel selection, parameter tuning, and scaling; complexity grows with number of support vectors in prediction cost.

Reversal

Replacing margin maximization with direct empirical risk minimization without margin (e.g., plain perceptron without regularization) reduces emphasis on robustness and can lead to overfitting or sensitivity to noise.

Boundary

Applies to supervised binary (and extended to multiclass) classification and regression (SVR) problems; less suitable without kernelization for highly structured outputs, or when probabilistic outputs are required without calibration; requires numerical feature representations.

Semantic Tension

Sometimes conflated with kernel methods generally or with large-margin ideas; SVM is specifically the maximum‑margin classifier solvable by convex quadratic programming (or equivalent formulations), while other kernelized learners trade the margin principle for different loss functions or probabilistic interpretations.

Synthesis

Support vector machines find hyperplanes that maximize class-separating margins, using support vectors to define sparse solutions and kernels to handle nonlinearity: a convex, margin-centered framework that delivers robust boundaries but requires careful kernel and regularization choices and does not natively provide calibrated probabilities.