Nu support vector regression
Algorithms AIMech:NuSupportVectorRegression DRAFT
Aliases: Nu-SVR, ν-SVR, Nu-SV regression
A supervised regression algorithm that jointly fits a regularized kernel expansion and a nonnegative constant epsilon-insensitive tube half-width, using nu to penalize the width relative to linearly penalized residual excesses.
Identity Notes & Concept Scope
The original symmetric constant-width nu-SVR in Scholkopf, Smola, Williamson and Bartlett (2000) is the selected identity. It differs from fixed-epsilon SVR parameterization, the paper’s supplied parametric-insensitivity extension, later learned-insensitivity/twin models, supervised nu-SVC and one-class support estimation. Compared all442 pinned AIO IDs, labels and scoped synonyms against all520 preceding native IDs, names and aliases, without a normalized collision. Full-text fixed-string whole-word searches in both checkouts included ignored and hidden files, expanded labels, acronyms, DOI and paper identifiers. Read all25 distinct matching text windows across88 locations: these are the existing one-class record context, its review/renderings and the preceding discovery lead, not supervised nu records. All125 semantic AIO rows and115 prior native name/definition rows match the previous reading; the additional one-class row and complete record were read. Thirty-one contextual full-record readings are retained with matching hashes. Absence is bounded to this snapshot and corpus.
Claims & Grounded Evidence
7 Claims
Mechanism
#learned-tube-objective
For n>=1 input/real-response pairs and a positive-semidefinite kernel with feature map Phi, fit f(x)=w.Phi(x)+b by minimizing (1/2)||w||^2+C*(nu*epsilon+sum_i(xi_i+xi_star_i)/n), with C>0, nu>=0, epsilon>=0 and nonnegative slacks. The constraints are f(x_i)-y_i<=epsilon+xi_i and y_i-f(x_i)<=epsilon+xi_star_i. Thus the symmetric, input-independent tube half-width is optimized jointly with the fit; C and the kernel still require specification.
Mechanism
#dual-and-recovery
Write beta_i=alpha_star_i-alpha_i. The dual maximizes sum_i beta_i*y_i-(1/2)*beta^T*K*beta, subject to sum_i beta_i=0, 0<=alpha_i,alpha_star_i<=C/n and sum_i(alpha_i+alpha_star_i)<=C*nu. Prediction is sum_i beta_i*k(x_i,x)+b. A free alpha_i gives f(x_i)-y_i=epsilon; a free alpha_star_i gives y_i-f(x_i)=epsilon. These equations can recover b and epsilon when adequate free coefficients exist. Otherwise the original KKT inequalities must be respected; neither an interior coefficient nor unique b, epsilon or dual coefficients is guaranteed.
Scope
#conditional-nu-counts
At an exact optimum with epsilon>0, the fraction of training observations strictly outside the tube is at most nu, and the fraction with a nonzero expansion coefficient is at least nu. The two sides separately obey the corresponding nu/2 bounds relative to the entire sample, not relative to each side. These are inequalities, not exact finite counts or calibrated test error. For four zero inputs with responses -1,-1,1,1, C=1 and nu=1/2, an exact optimum has b=0, epsilon=1 and four free coefficients of magnitude1/8: every point is a support vector and none is outside. The paper’s asymptotic equality additionally uses iid data, conditional-density regularity and function-class arguments; it is not asserted here unconditionally.
Scope
#fixed-epsilon-correspondence
Fixing epsilon at a value learned by nu-SVR, while retaining the same C, gives an epsilon-SVR problem sharing that fitted primal optimum. This conditional correspondence does not make arbitrary independently selected epsilon and nu interchangeable. The original nu-SVR estimates one constant tube half-width; the asymmetric quantile losses and supplied heteroscedastic basis functions described in this paper change the objective or constraints. Later learned parametric-insensitive methods remain pending separate full-method assessment and are not covered by this source claim.
Limitation
#degenerate-width-and-endpoints
A positive nu need not produce a positive epsilon. With all inputs and responses zero, zero coefficients, b=0 and epsilon=0 attain the exact minimum, so the positive-width support-count lower bound cannot be applied. At nu=0 an unpenalized sufficiently wide tube permits a constant fit. For nu>1, epsilon=0 reduces the loss to absolute residuals. At nu=1, epsilon and b can be nonunique: with two zero inputs, responses -1 and1 and C=1, (b,epsilon)=(0,0),(1/2,0),(0,1) all attain objective1. The fitted function or accepted tube can therefore vary between optimal solutions.
Limitation
#robustness-and-parameter-limits
The linear penalty outside the tube yields resistance to local response changes that keep those observations strictly outside on the same side; an existing optimum remains valid after matching slack changes. This is not resistance to arbitrary input leverage or unlimited contamination. The paper’s proposed asymptotically optimal nu calculation assumes a specified additive-noise scale family and convergence of the regression estimator, and is justified in a location-estimation model. The authors explicitly identify these restrictions; it is not a universal automatic nu selector.
Limitation
#evaluation-scope
The authors use LOQO for selected noisy-sinc experiments and a Boston-housing regression study. Sinc risk measures integrated absolute deviation from the noiseless target; the housing study reports response MSE over 100 random splits of 506 observations into 481 training and 25 test observations. Table 3 reports sweeps over nu and fixed epsilon, with other settings held fixed. The text compares nu-SVR with the best fixed-epsilon test result after evaluation; this is not nested parameter validation or proof of universal superiority. The supplied-noise-shape experiment studies a separate parametric-insensitivity extension. No numerical performance advantage is curated.
⚠️ Curated Limitations & Negative Results
- Finite nu bounds are conditional inequalities, not exact sparsity, calibrated test error or a universal parameter choice.
- All39 original journal pages were read; the conference precursor, later decomposition methods and external proof dependencies were not fully audited.
- Exact rational feasibility and KKT certificates were evaluated without running a solver, training a model or reproducing the historical benchmarks.
- The record remains a native DRAFT. Parametric-insensitivity and later solver/extension identities remain separate discovery work.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:RegressionAnalysis
Regression Analysis
|
related_match | AIO RegressionAnalysis identifies a supervised estimation task, while this native record specifies one fitting algorithm for real-valued responses. A related match captures its task without asserting algorithm/task equivalence. LinearRegression, RidgeRegression and LeastsquaresAnalysis do not identify the kernel, learned tube or piecewise-linear loss. |
AIO:SupportVectorMachine
Support Vector Machine
|
related_match | The source derives this supervised kernel optimization from support-vector methods, but the pinned AIO target describes a network with input, hidden and output parts. This native identity is an optimization algorithm, so the network abstraction is related rather than an asserted exact or broader class. KernelLayer names a component, and generic Network or TrainingStrategy also does not identify the objective. Twelve annotated AIO CLASS nodes and their outgoing edges were inspected. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
scholkopf2000
|
Bernhard Schölkopf, Alex J. Smola, Robert C. Williamson, Peter L. Bartlett
|
primary paper
full_text
|
ID: https://doi.org/10.1162/089976600300015565
sha256:57eb6d4db2e8…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-12T11:18:19Z
Updated: 2026-10-12T11:25:45Z
Source File:
data/curated/NuSupportVectorRegression.yaml