L2-loss support vector data description
Algorithms AIMech:L2LossSupportVectorDataDescription DRAFT
Aliases: L2-loss SVDD, Squared-hinge-loss support vector data description
A support-vector data-description algorithm that learns a feature-space sphere using squared-hinge penalties on squared-distance violations, yielding a unique optimal center, squared radius and slack vector for a fixed positive penalty and feature map.
Identity Notes & Concept Scope
This is the target-only L2-loss extension in Chang, Lee and Lin Section 4.1. L2 denotes squared slack loss, not merely the Euclidean geometry already used by standard SVDD. It changes the dual penalty and zero-radius threshold and is distinct from linear-slack SVDD, negative-example SVDD, origin-separating one-class SVM and support-vector clustering. 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
4 Claims
Mechanism
#squared-slack-objective
For C>0, minimize t+C*sum_i xi_i² subject to t>=0 and ||Phi(x_i)-a||²<=t+xi_i. Explicit nonnegativity constraints on xi are redundant at an optimum: a negative slack can be replaced by zero to improve the objective. This convex formulation has strong duality and a unique optimal center, squared radius and slack vector. Its squared violation penalty changes the method, even though both SVDD variants use squared Euclidean distances.
Mechanism
#zero-radius-threshold
First consider the zero-radius auxiliary problem minimizing sum_i ||Phi(x_i)-a||^4, with optimal squared distances xi_i*. When their sum is positive, define C*=1/(2*sum_i xi_i*). For 0<C<=C*, the optimal squared radius is zero and the auxiliary center is optimal; for C>C* the radius is positive. This threshold depends on the data and feature map and is not the standard SVDD threshold 1/N. The auxiliary center need not be the sample mean. If all training features coincide, the optimal radius and slacks are zero at their common feature and the denominator formula is inapplicable.
Mechanism
#positive-radius-dual
For C>C*, maximize sum_i alpha_i*K_ii-alpha^T*K*alpha-sum_i alpha_i²/(4*C) subject to alpha_i>=0 and sum_i alpha_i=1. The additional quadratic term makes the maximization objective strictly concave; there is no standard SVDD upper bound alpha_i<=C, though the simplex implies alpha_i<=1. Recover a=sum_i alpha_i*Phi(x_i), xi_i=alpha_i/(2*C), and t=||Phi(x_i)-a||²-alpha_i/(2*C) using any positive coefficient. This equality-simplex recovery is scoped to the positive-radius regime and must not be applied indiscriminately below C*.
Limitation
#prediction-and-evidence-limits
Prediction compares feature-space squared distance to the learned squared radius, using the kernel expansion when appropriate. A zero-radius fit accepts only queries mapped exactly to the center. The reassessment supplies mathematical derivations rather than a benchmark comparison establishing superiority of squared-hinge SVDD over linear-slack SVDD. Standard SVDD support-count and box-constraint error arguments do not transfer unchanged to this loss. For the symmetric linear-kernel sample {-2,-1,1,2}, C*=1/20; at C=1/10 the solution has center 0, t=3/2 and objective 11/4, as verified by exact KKT and primal-dual calculations.
⚠️ 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 21 manuscript pages, including proofs, were visually read. The differently formatted author copy and implementation archive were not inspected in full. The software index alone does not verify an implementation of this specific L2-loss SVDD extension.
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
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
chang-revisit
|
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/L2LossSupportVectorDataDescription.yaml