Home / Curated / Algorithms / Support vector data description

Support vector data description

Algorithms AIMech:SupportVectorDataDescription DRAFT

Aliases: SVDD, Support vector domain description, L1-loss SVDD

A one-class support-estimation algorithm that learns a sphere in an input or kernel feature space by minimizing squared radius plus linearly penalized nonnegative distance-constraint violations, then accepts a query when its feature lies within or on the learned sphere.

Identity Notes & Concept Scope This record covers the target-only linear-slack method called support vector domain description in 1999 and support vector data description in 2004, with later qualifications to its optimization. L1 refers to the slack loss, not an L1 norm on the sphere center. The negative-example extension and squared-hinge-loss variant change the optimization and are not exact aliases. Sphere support estimation does not include SVC graph-component labeling or the general origin-separating one-class hyperplane formulation. Compared all 442 pinned AIO IDs, labels and scoped synonyms with all 502 preceding native IDs, names and aliases. No normalized candidate identity collision was found. Full fixed-string whole-word searches included ignored and hidden files in both checkouts, expanded names, acronyms and source identifiers. All 15 matches per checkout, eight distinct texts at 30 locations, were inspected as source queries, search terms and pending leads in the preceding discovery ledger, not existing concepts. Ten annotated AIO nodes and outgoing edges, their generated records, and three native neighbors were inspected or checked against unchanged complete readings. Four SAT Novelty-family records were inspected only for names, definitions and identity notes: those flip heuristics are unrelated. Absence is bounded to this pin and inspected corpus.

Claims & Grounded Evidence

6 Claims

Mechanism #convex-formulation

For N target observations and C>0, use squared radius t>=0, center a and slacks xi_i>=0 to minimize t+C*sum_i xi_i subject to ||Phi(x_i)-a||²<=t+xi_i. This is a convex reformulation of the radius-variable problem and satisfies a strict-feasibility condition supporting strong duality. The objective penalizes violation magnitudes, not an exact count of rejected examples. The optimal center is unique, but the radius and slacks can have multiple optima.

supports Source: Support Vector Data Description · Locator: Section 2.1, printed pp.47-49 / PDF pp.3-5, Equations 1-15
supports Source: A Revisit to Support Vector Data Description · Locator: Sections 2.2-3.1, PDF pp.4-7, Equation 7, Theorem 2 and Corollary 1; Appendix A, pp.15-16

Mechanism #dual-and-small-penalty

For C>1/N, the usual dual maximizes sum_i alpha_i*K_ii-alpha^T*K*alpha subject to sum_i alpha_i=1 and 0<=alpha_i<=C, and a=sum_i alpha_i*Phi(x_i). For 0<C<1/N this equality-simplex dual is infeasible even though the primal is feasible: its optimum has t=0 and the feature-space sample mean as center. At C=1/N the same zero-radius solution is optimal, but additional positive-radius optima may exist. Dividing the original radius stationarity equation by R silently omits R=0. For C>1 the method reduces to the minimum sphere enclosing all training features.

supports Source: A Revisit to Support Vector Data Description · Locator: Sections 2.3 and 3.2, PDF pp.5-11, Theorems 4 and 6, Equations 9-26; Appendices B and D, pp.16-17

Mechanism #radius-and-prediction

For C>1/N and an optimal dual solution, squared distances d_i² to the center bound an optimal t by max_{alpha_i<C} d_i² <= t <= min_{alpha_i>0} d_i². A free coefficient 0<alpha_i<C fixes the radius; otherwise choosing the interval midpoint is an implementation convention. Bounded coefficients need not be strictly outside the sphere. Accept a query when K(x,x)-2*sum_i alpha_i*K(x_i,x)+alpha^T*K*alpha<=t. For observations {-2,-1,1,2} with a linear kernel and C=1/2, a=0 and every t in [1,4] achieves objective 4; this illustrates different acceptance regions at the same optimum.

supports Source: Support Vector Data Description · Locator: Section 2.1, printed p.49 / PDF p.5, Equations 14-15
supports Source: A Revisit to Support Vector Data Description · Locator: Section 3.2, PDF pp.9-11, Theorem 5 and Equations 20-23; Appendix A, p.16, Equation 41; Appendix C, p.17

Mechanism #gaussian-and-boundary-scope

A Gaussian kernel exp(-||x-y||²/s²) gives a weighted kernel-score acceptance region without explicitly constructing features. It can produce a non-spherical region in input space. The narrow-width equal-weight limit assumes distinct training inputs; the exact constant-kernel limit is degenerate and does not select a unique coefficient vector. Comparisons with one-class hyperplane support estimation require the stated kernel/normalization conditions rather than general method identity. The boundary estimate is not by itself a normalized density model or a component-labeling rule.

supports Source: Support vector domain description · Locator: Section 3, printed pp.1193-1194 / PDF pp.3-4
supports Source: Support Vector Data Description · Locator: Sections 2.1 and 2.3, printed pp.50-54 / PDF pp.6-10, Equations 16-21 and 30-35; curator constant-Gram qualification

Limitation #support-count-and-error

The papers use support-vector fractions to guide target-error and kernel-width choices. The 2004 account distinguishes essential from nonessential vectors and gives a leave-one-out bound/estimate; the raw support fraction is not universally equal to observed leave-one-out rejection or a calibrated future error rate. With two identical training features and C=2, a valid two-support-vector representation has radius zero, while either leave-one-out fit accepts the omitted duplicate. For the nonnegative equality-simplex coefficients the fraction at the upper bound is at most 1/(N*C); bounded coefficients can be on the boundary, and this count bound cannot be transferred unchanged to signed negative-example coefficients.

supports Source: Support vector domain description · Locator: Section 4, printed pp.1194-1195 / PDF pp.4-5, Equation 11
supports Source: Support Vector Data Description · Locator: Section 2.4, printed pp.55-56 / PDF pp.11-12, Equation 36; curator duplicate-feature and coefficient-sum calculations

Limitation #empirical-and-extension-boundaries

The published tests depend on target/negative sampling, feature dimension, preprocessing, kernel width and target-rejection settings. The 2004 paper reports failures when a pump training set covers too little operating range and cautions that overlapping negative examples can worsen its separate negative-example extension. Its reverse distance constraints are not the convex target-only formulation above. These studies do not establish universal performance superiority or immunity to distribution changes.

supports Source: Support vector domain description · Locator: Sections 5-6 and Discussion, printed pp.1195-1199 / PDF pp.5-9
supports Source: Support Vector Data Description · Locator: Section 2.2, printed pp.51-52 / PDF pp.7-8; Sections 3-5, pp.57-65 / PDF pp.13-21, Tables 1-4 and Figures 4-7

⚠️ Curated Limitations & Negative Results

Kernel, preprocessing, regularization and decision-threshold conventions affect the acceptance region. The method does not guarantee a calibrated future outlier probability, discovery of all abnormal cases, or optimal cluster labels.
The later mathematical reassessment is an author-hosted technical manuscript with conflicting author-index year labels. Exact URL and hash preserve inspected evidence; journal publication and version equivalence are not asserted.
Exact finite calculations check mathematical edge cases. No historical implementation, training run or benchmark was executed.
All 9, 22 and 21 pages of the three cited PDFs were visually read. Equations with degenerate-radius assumptions, categorical multiplier descriptions and unqualified support-count interpretations are explicitly qualified. No new numerical benchmark result is asserted.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:BinaryClassification` Binary Classification	related_match	AIO:BinaryClassification is related through the accept-versus-reject prediction task. Its Classification parent is under SupervisedLearning, whereas this particular algorithm learns a target-class description without requiring representative negative training examples. This is a related task mapping, not algorithm-task equivalence, an is-a assertion or a claim of two-class supervision. Inspected BinaryClassification, Classification, SupervisedLearning, UnsupervisedLearning, MachineLearningTask, SupportVectorMachine, Model, Network, KernelLayer and UnsupervisedClustering, including definitions and outgoing edges. The supervised category-separating network definition of SupportVectorMachine does not identify a sphere support estimator. The clustering task assigns groups, while this method supplies an acceptance region; generic models, networks and kernel layers do not identify the algorithm. All 442 class labels and 153 semantic matches retain byte-verified prior readings; expanded description/hinge/enclosing/classification searches added no rows. SupportVectorClustering adds graph labeling; DBSCAN uses neighborhood density connectivity; KernelPCA extracts variance directions. These contexts do not receive renewed scientific approval.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`tax1999`	Support vector domain description David M. J. Tax, Robert P. W. Duin	primary paper full_text	sha256:3a5748fa74d9…
`tax2004`	Support Vector Data Description David M. J. Tax, Robert P. W. Duin	primary paper full_text	sha256:fe7e0705e66b…
`chang-revisit`	A Revisit to Support Vector Data Description Wei-Cheng Chang, Ching-Pei Lee, Chih-Jen Lin	primary paper full_text	sha256:7618a0ff3b92…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-12T07:36:18Z

Updated: 2026-10-12T07:36:18Z

Source File: data/curated/SupportVectorDataDescription.yaml