Synthetic Minority Over-sampling Technique
Preprocessing Methods AIMech:SyntheticMinorityOversamplingTechnique DRAFT
Aliases: SMOTE
A minority-class oversampling method that generates synthetic labeled feature vectors by interpolating between an observed minority example and selected neighbors from the same class.
Identity Notes & Concept Scope
This record denotes the original continuous-feature SMOTE method. It generates training examples rather than fitting a classifier. It differs from replacement sampling, majority undersampling, cross-class input/label mixing in MixUp, and hidden-representation mixing. SMOTE-NC and the proposed nominal-only SMOTE-N extension require separate feature-distance and synthesis rules. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 273 preceding native IDs/names/aliases. Full-checkout fixed-string searches included ignored and hidden files in worktree and main, returning 18 and 18 lines respectively. Every text hit was a prior candidate lead. Eleven exact AIO CLASS nodes with scoped synonyms/comments and outgoing edges and five native neighboring mechanisms were inspected. Absence is bounded to this snapshot and semantic neighborhood.
Claims & Grounded Evidence
1 Claim
Mechanism
#method
For each selected minority-class training example, find a configured number k of nearest neighbors within the minority class and randomly choose neighbors according to the requested oversampling amount. In the prose and Table1 formulation, construct a synthetic vector as the original vector plus a random fraction between zero and one of the neighbor-minus-original vector, retaining the minority-class label. The paper uses k=5 in its implementation, but k and the oversampling amount are parameters. Its pseudocode treats percentages below100 by selecting a random subset of minority examples; higher amounts are expressed as integral multiples of100. Majority-class undersampling is a separately described operation that may be combined with SMOTE, not a required step of synthetic generation.
⚠️ Curated Limitations & Negative Results
- The Section4.2 prose and Table1 describe one random scalar per synthetic vector, placing it on a line segment. Pseudocode line21 draws gap inside the attribute loop; independent per-coordinate draws generally produce a point in the coordinatewise rectangle rather than on that segment. The record follows the stated interpolation mechanism and does not certify these descriptions as identical executable code.
- Interpolation requires a meaningful continuous feature space and suitable neighbor distances. It does not guarantee a new distinct vector, a plausible observation, or a point inside the true minority-class support; identical endpoints and overlap with majority regions can defeat those interpretations.
- The original paper’s nominal-continuous and nominal-only extensions are distinct from this core method. Sampling percentages and tie/neighbor conventions need implementation choices; the pseudocode is not a complete contract for every modern library.
- The source discusses possible overlap with majority-class space and results that vary by dataset and learner. Synthetic labels are an inductive assumption, not new observed ground truth, and oversampling does not universally improve minority detection or calibrated probabilities. No comparative performance result is asserted.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:DataAugmentation
Data Augmentation
|
broad_match | SMOTE constructs additional minority-class training vectors by transforming existing examples through within-class interpolation. The pinned augmentation class is broader than this particular continuous-feature synthesis procedure; this is a local-to-broader mapping, not identity or an upstream ontology edit. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Nitesh V. Chawla, Kevin W. Bowyer, Lawrence O. Hall, W. Philip Kegelmeyer
|
primary paper
full_text
|
sha256:8eece9837dbc…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T12:27:02Z
Updated: 2026-10-11T12:27:02Z