Free scoping read. A written assessment of your manuscript in two business days, at no cost.

Start one
Contents

Academy · Research Data, Analytics & AI · Support Vector Machines

Support Vector Machines: The Maximum-Margin Principle, Support Vectors and Kernel Functions

A support vector machine does not attempt to fit the training data as closely as possible. Instead, it identifies the boundary with the greatest separation from the nearest points on both sides, a property that is associated with good generalisation to unseen data.

Pubrica AcademyWritten by a data-analytics reviewer Reviewed bya biostatistician Published10 October 2026 Last checked10 October 2026 Reading time8 minutes Sources1 peer-reviewed

A support vector machine (SVM) is a supervised machine-learning algorithm that classifies data by identifying the hyperplane that separates classes with the maximum possible margin, defined as the largest distance to the nearest data points of either class. Through kernel functions, the approach can be extended to data that are not linearly separable.

Learning objectives

  • The meaning of the “maximum margin” principle, and why it is the defining concept of support vector machines.
  • The definition of “support vectors”, and why most of the training data does not influence the final model.
  • The function of the kernel trick, and how it allows support vector machines to model non-linear boundaries.
  • The types of task for which a support vector machine is suited, relative to other classifiers.

Common misconceptions

  • “A support vector machine identifies any boundary that separates the classes.” A support vector machine identifies the boundary that maximises the margin to the nearest points of each class; among the infinite number of lines that could separate linearly separable data, only one has the maximum margin, and that is the boundary the algorithm selects.
  • “Every training example contributes equally to the final SVM boundary.” Only the data points closest to the boundary (the support vectors) determine its position; points distant from the boundary could be removed without any change to the trained model.
  • “Support vector machines can model only linear boundaries.” With a kernel function, a support vector machine operates implicitly in a higher-dimensional transformed space, in which a linear boundary corresponds to a non-linear boundary in the original feature space, without explicitly computing the transformed coordinates.

The maximum-margin principle

Cortes and Vapnik, in the 1995 paper that introduced the modern soft-margin support vector machine, define the problem directly: given data from two classes, the task is to find the hyperplane that separates them with the largest possible margin, that is, the distance from the hyperplane to the nearest point of either class1. This is a specific, well-defined optimisation target, distinct from finding any line that happens to separate the training data; among all boundaries that would classify the training set correctly, the SVM selects the one with the widest margin on either side, which is the property that confers good generalisation to new data in practice.

Support vectors and the limited influence of most training data

The position and orientation of the hyperplane are determined entirely by the training points that lie closest to it, the support vectors from which the algorithm takes its name. Cortes and Vapnik's formulation shows that the optimal separating hyperplane can be expressed solely in terms of these support vectors; points further from the boundary do not appear in that expression, and could therefore be removed from the training set without changing the fitted model1. The soft-margin extension in their 1995 paper addresses a related issue: real data are often not perfectly separable, so the method permits some points to violate the margin, subject to a penalty parameter, accepting a small amount of training error in exchange for a more robust boundary.

The kernel trick

Many real classification problems are not linearly separable in their original feature space. The kernel trick, incorporated in the SVM framework that Cortes and Vapnik describe, replaces the direct dot-product computation between data points with a kernel function that computes the value of that dot product in a higher-dimensional, implicitly defined transformed space, without explicitly constructing the coordinates in that space1. A boundary that is a flat hyperplane in the transformed space can correspond to a complex, non-linear boundary in the original feature space; in this way an SVM with a non-linear kernel (such as a radial basis function kernel) can classify data that a linear boundary could not separate.

Suitability of support vector machines relative to other classifiers

SVMs tend to perform well on classification tasks with a clear margin of separation and a moderate number of features relative to the number of training examples, and they have an established record in biomedical classification tasks of this kind. They are less well suited to problems with very large training sets (training time and memory scale poorly with dataset size in the classic formulation) or to tasks in which the relevant structure is sequential or spatial, for which architectures designed for that structure are typically preferred.

Key points

Key points

A support vector machine identifies the hyperplane with the maximum margin between classes, a boundary determined only by the nearest points (the support vectors). The kernel trick extends this maximum-margin principle to data that are not linearly separable by operating implicitly in a higher-dimensional space.

Practical application

Practical application

  1. The number of features and the sample size should be assessed for suitability for a support vector machine. Support vector machines tend to perform well with moderate numbers of features and examples; for very large training sets, the computational cost should be considered before the classic formulation is adopted.
  2. The kernel should be selected and justified rather than adopted by default. Authors should state whether a linear or non-linear (e.g. radial basis function) kernel was used and why, because this choice materially affects the boundaries that the model can represent.
  3. The margin-penalty (regularisation) parameter should be reported. The soft-margin penalty parameter controls the trade-off between margin width and training error; the method by which it was chosen (e.g. cross-validation) should be reported rather than left as an unstated default.

Frequently asked questions

No. Both can produce a linear decision boundary, but they are fitted by different optimisation objectives (logistic regression maximises likelihood under a probabilistic model, whereas an SVM maximises the margin between classes), and they can behave differently, particularly near the decision boundary and when kernels are used.

Not directly; the base algorithm outputs a class decision according to the side of the margin on which a point falls. Probability estimates can be obtained through additional calibration methods applied to the base SVM output.

It refers to the 1995 Cortes and Vapnik extension that permits some training points to fall inside the margin or on its wrong side, subject to a penalty parameter, so that the method can still identify a useful boundary for data that are not perfectly linearly separable.

References

  1. Cortes C, Vapnik V. Support-vector networks. Mach Learn. 1995;20:273-297. read 10 October 2026. View on Machine Learning (Springer)

Every source above is a peer-reviewed article, a publisher’s record of one, or a standard issued by the body that owns it. Each carries the date it was last read. Where a standard moves, this article is revised rather than left up with a note.

How to cite this article

Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.

Pubrica Academy. (2026, October 10). Support Vector Machines: The Maximum-Margin Principle, Support Vectors and Kernel Functions. Pubrica. https://pubrica.com/academy/concepts-definitions/support-vector-machines/

This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.

SVM reporting checklist for a manuscript

This checklist applies when a support vector machine model is described in the methods section.

[ ] Kernel type stated (linear, polynomial, radial basis function, etc.) and why chosen
[ ] Margin-penalty (regularisation) parameter and how it was selected
[ ] Feature scaling/normalisation approach stated
[ ] Train/validation/test split or cross-validation scheme stated
[ ] Class balance of training data stated

Pubrica AcademyWritten and maintained by Pubrica’s subject matter experts. Articles are revised when the underlying standards change, and each article states the date on which it was last reviewed.

How Pubrica can help

Pubrica’s data analytics team develops, validates and reports support vector machine and other classification models for research manuscripts, with the choice of kernel, parameter selection and validation reported to the standard expected by statistical reviewers.

Find something else

Search the Pubrica Academy

Access is free and requires no account. Enter a topic, or open the full library to filter by category, sub-topic or resource type.

Pubrica can support the reporting of classification models so that the methods section withstands peer review.

A scoping assessment is free and takes two business days. It provides a written assessment of what the work requires and which specialist would undertake it.

Get a free quote

Regulations

Check your project against this rule

  1. 1Your work
  2. 2Your details
  3. 3Send
What do you need?

Material you send is held under signed confidentiality, with access limited to the assigned team and encrypted transfer. It is not used to train anything and is not passed to third-party tools. Ask for the agreement before you send anything you would rather cover first.