RAPID support vector data description sampling
Preprocessing Methods AIMech:RapidSupportVectorDataDescriptionSampling DRAFT
Aliases: RAPID, Reducing samples by pruning of inlier densities
A density-based training-subset selection method for support vector data description that prefilters low-density observations and greedily removes high-density inliers while testing a retained-density threshold on the original inlier set.
Identity Notes & Concept Scope
RAPID denotes the Englhardt-Trittenbach-Kottke-Sick-Böhm density-pruning procedure, expanded as Reducing sAmples by Pruning of Inlier Densities. The intended candidate-set rule, printed pseudocode and versioned implementation conventions are distinguished below. It is neither an SVDD prediction rule nor a generic random-sampling algorithm. Compared all 442 pinned AIO IDs, labels and scoped synonyms with all 518 preceding native IDs, names and aliases. No normalized identity collision was found. Fixed-string whole-word searches included ignored and hidden files in both checkouts, expanded names, RAPID, the publication title, DOI and arXiv identifier. All 126 distinct matched-text windows were read: these are prior discovery/review references, pending RAPID leads and ordinary uses of rapid, not an existing RAPID concept. All 125 semantic AIO rows and 114 native name/definition rows were read or reused after exact equality checks. Annotated AIO nodes and outgoing edges, and 29 contextual YAMLs with full readings, distinguish preprocessing, prediction tasks, neural components and neighboring subset-selection methods. Absence is bounded to this snapshot and maintained corpus.
Claims & Grounded Evidence
8 Claims
Mechanism
#density-prefilter
For a nonempty finite observation set X and a unit-diagonal kernel k with k(x,x)=1, compute the unnormalized empirical density d_X(x)=sum_{z in X} k(x,z). An expected-outlier parameter selects a low-density threshold; keep I={x in X:d_X(x)>=theta_pre}, discard the other observations and recompute densities using only I as kernel centers. Subsequent subset selection prepares data for hard SVDD with C=1. This prefilter-plus-hard-fit procedure is not identical to fitting arbitrary-C soft SVDD on the original data. The paper leaves quantile index rounding, endpoints and ties under-specified; exact requested rejection fractions are not assured.
Mechanism
#greedy-pruning-rule
The stated intended rule starts with S=I and, while |S|>1, selects a highest-density retained observation r, and subtracts its kernel contribution at every original inlier. For the candidate T=S minus {r}, define theta_T=min_{z in T} d_T(z). Accept the removal only if d_T(x)>=theta_T for every x in I; otherwise stop and return the preceding S. If S reaches a singleton, return it without testing an empty candidate. This intended-rule stopping convention is distinct from the stricter public-code floor described below. The returned observations are used to train SVDD. The rule is a greedy density-surrogate selection, not repeated SVDD optimization, a globally minimum-cardinality subset, or a certified global solution of the mixed-integer problem.
Limitation
#optimization-and-guarantee-scope
The source SOP minimizes the range theta_max-theta_min of empirical densities over the original inlier set, subject to nonempty binary subset selection and a selected minimum-density witness. Minimum cardinality is a tie-breaker between SOP optima. Theorem 1 establishes a finite-inlier label statement for exactly uniform densities, not a quantitative bound on a continuous decision boundary under approximate uniformity. Theorem 2 uses separated density extrema and a small-bandwidth limiting argument, while its displayed comparison treats old extrema as if they remained extrema after removal. Those steps do not establish an unconditional finite-data SVDD boundary-preservation guarantee for greedy pruning.
Limitation
#finite-density-boundary-distinction
Unweighted kernel-density level sets and fitted finite-sample SVDD decision regions are generally different. An exact curator example uses k(x,y)=2^(-(x-y)^2) on points 0,1,4. Positive hard-SVDD coefficients (1205376,1198391,1802240)/4206007 give the same weighted boundary score 1804599/4206007 at all three nonstationary boundary crossings. Their unweighted densities are respectively 98305/65536, 769/512 and 65665/65536, so no single unweighted-density threshold has all three as boundary points. This challenges an unqualified global identity, not the trivial fact that two rules can accept the same finite inlier set. The cited Vert-Vert results concern asymptotic density-level estimation under distribution, regularization and bandwidth conditions; they do not supply that finite global identity.
Limitation
#printed-pseudocode-discrepancy
Printed Algorithm 2 subtracts r from the density sum but computes the minimum over the old S before deleting r. This differs from the candidate-set rule in the prose and inspected Julia code. For the specified Gaussian kernel 2^(-(x-y)^2), with retained inliers 0,1,2, removing the middle point leaves densities (17/16,1,17/16). The old-set minimum 1 accepts the removal, whereas the candidate-set minimum 17/16 rejects it because the removed middle observation is still an original inlier. The literal printed flow can continue to one point for this nonidentical dataset. This is exact pseudocode arithmetic after retaining all three inliers, not an execution of the public code or its percentage-prefilter convention.
Limitation
#versioned-code-conventions
At the inspected public commit, RAPID removes r from the candidate mask before the minimum test and reconstructs the final mask from accepted removals, undoing a failed candidate. For m initial inliers, its loop starts at i=1 and continues while i<m-2, permitting at most m-3 accepted removals when m>=3; smaller sets are unchanged. The percentage threshold clamps floor(p_out*N) to the one-based index range and averages adjacent sorted densities. With a unique minimum it can exclude one observation at p_out=0; equal-density ties can retain all observations. The source also offers a ground-truth mask and a mode deriving p_out from supplied labels, so all package configurations are not unsupervised. These are source-code observations, not verified historical experimental behavior.
Limitation
#kernel-and-experiment-version
The paper calls exp(-gamma*||x-y||) Gaussian in its displayed formula, omitting the squared norm. The inspected sampling code delegates to MLKernels.GaussianKernel; the evaluation-pinned MLKernels v0.4.0 implements exp(-gamma*||x-y||^2). These formulas must not be silently conflated. Although the MLKernels tree matches the evaluation manifest, its OneClassSampling entry points to an author-local development directory without a revision or tree hash. The inspected public sampling commit therefore cannot be identified as the exact experimental version from that manifest.
Limitation
#cost-and-evaluation-boundaries
The source gives O(N^2) sampling work under unit-cost kernel evaluation, and its optional Gram matrix and inspected KDE cache use quadratic storage. Input dimension, parameter selection and the downstream SVDD fit add costs. The experiments compare selected synthetic mixtures and real outlier datasets with specified bandwidth heuristics and prefiltering; the main setup supplies the true outlier percentage. Some costly competitors are excluded and the full-data baseline is limited to smaller datasets. These results do not establish universal speed, accuracy, compression or preserved generalization. The method still requires a kernel and an outlier-setting convention; no numerical performance advantage is curated.
⚠️ Curated Limitations & Negative Results
- A density-surrogate feasibility check is not an unconditional guarantee of finite SVDD boundary or prediction equivalence.
- Printed pseudocode, public code and percentage-prefilter conventions differ. The exact historical sampling package revision is not pinned by the inspected evaluation manifest.
- All 27 final RAPID PDF pages were visually read; source files and scoped consistency-theory passages were inspected. Exact rational formula checks were executed, but no Julia code, optimizer, model training or benchmark was run.
- This remains a native DRAFT. Competing boundary selectors, active-learning parameter methods and bandwidth heuristics require their own primary-method assessments.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:Preprocessing
Preprocessing
|
broad_match | AIO:Preprocessing is broader: RAPID transforms the available observations into a selected training subset before an SVDD model is fitted. This mapping states the intended role of the procedure, not a guarantee of improved data quality or classifier accuracy. DataPreparation and TrainingStrategy are generic neighboring scopes, not exact names for this density-pruning method. BinaryClassification is the downstream accept/reject task rather than the selection algorithm. SupportVectorMachine describes a supervised category-separating network; KernelLayer is a neural component. SelectionAndSamplingBias is a bias condition, not a sample-selection procedure. Native SVDD fits a sphere; CSVDD grows a core using repeated SVDD fits, whereas RAPID prunes a density surrogate without those fits. CNNDD and CPDD use different distance-based consistency or coverage rules. |
Evidence Sources & Primary Literature
5 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
englhardt2022
|
Adrian Englhardt, Holger Trittenbach, Daniel Kottke, Bernhard Sick, Klemens Böhm
|
primary paper
full_text
|
ID: https://doi.org/10.1007/s10994-022-06149-0
sha256:1e926feb7ef7…
|
rapid-code
|
Adrian Englhardt
|
official documentation
documentation
|
sha256:8a13216d478d…
|
evaluation-manifest
|
Adrian Englhardt
|
official documentation
documentation
|
sha256:f165f86c8013…
|
mlkernels
|
Tim Thatcher
|
official documentation
documentation
|
sha256:407693f7b5d9…
|
vert2006
|
Régis Vert, Jean-Philippe Vert
|
primary paper
full_text
|
sha256:e1bf86aad7a2…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-12T10:41:52Z
Updated: 2026-10-12T10:49:58Z