Home / Curated / Preprocessing Methods / SMOTE-NC

SMOTE-NC

Preprocessing Methods AIMech:SMOTENC DRAFT

The original mixed nominal/continuous extension of SMOTE that synthesizes minority-class training examples by finding minority neighbors with a continuous-distance and nominal-mismatch rule, interpolating their continuous coordinates and assigning nominal coordinates through neighbor votes.

Identity Notes & Concept Scope This record denotes the original mixed nominal/continuous SMOTE-NC preprocessing procedure, distinct from the existing continuous-feature SMOTE record and from the separately proposed nominal-only SMOTE-N. Fresh inventory of 739 maintained and 442 pinned AIO identities at merged base 50d9a115a61f9a8d40db79fef421ce23c3409bd9 found no normalized identity collision. Four ignored/hidden-inclusive checkout searches yielded 280/283/787/829 occurrences and ten complete distinct strings read by root, consisting of existing leads and continuous-SMOTE context or its historical/derived representations. Historical research inspected 47 selected definitions, four full neighboring records and thirteen exact AIO CLASS nodes with all 49 incident edges. Minority-only median standard deviations and nominal-neighbor voting distinguish this procedure; the original prose continuous interpolation interpretation and pseudocode discrepancy remain explicit. The local-to-broader DataAugmentation mapping is retained with its bounded rationale, not equivalence or an accuracy guarantee. No claim of exhaustive modern implementation coverage is made.

Claims & Grounded Evidence

6 Claims

Scope #mixed-feature-minority-synthesis

The original SMOTE-NC procedure extends continuous-feature SMOTE to training examples with both continuous and nominal features. It creates additional examples assigned to the minority class by using minority-class neighbors, continuous interpolation and nominal-feature votes. The procedure is training-data preprocessing rather than a fitted classifier. Its mixed-feature construction is distinct from the original continuous-only method and from the separately proposed nominal-only SMOTE-N construction.

supports Source: SMOTE: Synthetic Minority Over-sampling Technique · Locator: Section 6.1 introduction and steps 1–3, PDF pp. 28–29 / printed pp. 348–349; Section 4.2, PDF p. 8 / printed p. 328; Section 6.2 opening, PDF p. 29 / printed p. 349.

Mechanism #minority-continuous-standard-deviation-median

Compute the standard deviation of each continuous feature using the minority-class examples, then take the median M across those feature-wise standard deviations. This single scale supplies the nominal-mismatch penalty in neighbor distance. It is not a median of raw feature values, a median of variances, or a statistic pooled across all classes.

supports Source: SMOTE: Synthetic Minority Over-sampling Technique · Locator: Section 6.1 step 1 and the definition of Med in Table 6, PDF p. 28 / printed p. 348.

Mechanism #mixed-distance-with-squared-nominal-penalty

For minority examples x and y, the displayed distance rule is d(x,y)=sqrt(sum_continuous (x_j-y_j)^2 + M^2*sum_nominal 1[x_j!=y_j]). Each differing nominal coordinate contributes a separate M^2 inside the square root. The displayed example has two nominal mismatches and therefore two such terms. This rule does not assign numeric interpolation values to category labels and is not the value-difference construction discussed for SMOTE-N.

supports Source: SMOTE: Synthetic Minority Over-sampling Technique · Locator: Section 6.1 step 2 and Table 6, PDF p. 28 / printed p. 348; Section 6.2 opening, PDF p. 29 / printed p. 349. The mixed-distance example is captioned Table 6, despite the preceding prose reference to Table 2.

Mechanism #minority-neighbors-and-generation-parameters

For a minority-class seed, identify k nearest neighbors among minority examples using the mixed distance, then apply the earlier SMOTE generation procedure with randomly selected members of that neighbor set and a requested oversampling amount. The earlier implementation uses k=5, while its algorithm takes k and the oversampling percentage as inputs. This does not define five as a universal requirement or fully settle ties, self-inclusion and neighbor-reuse conventions.

supports Source: SMOTE: Synthetic Minority Over-sampling Technique · Locator: Section 6.1 steps 2–3, PDF pp. 28–29 / printed pp. 348–349; Section 4.2 prose and Algorithm SMOTE inputs and lines 13–18, PDF pp. 8–9 / printed pp. 328–329.

Mechanism #shared-scalar-continuous-interpolation

Following the earlier prose and Table 1 formulation, for the seed x and one selected minority neighbor y, generate the continuous subvector as x_cont+r*(y_cont-x_cont), using one random scalar r between zero and one for that synthetic example. The continuous subvector therefore lies on their continuous-coordinate line segment. This record explicitly follows that shared-scalar description: the printed pseudocode instead places the gap draw inside its coordinate loop, an unresolved source inconsistency.

supports Source: SMOTE: Synthetic Minority Over-sampling Technique · Locator: Section 6.1 step 3, PDF p. 29 / printed p. 349, referring to Section 4.2 prose on PDF p. 8 / printed p. 328 and Table 1 on PDF p. 10 / printed p. 330; compare Algorithm SMOTE lines 19–22, PDF p. 9 / printed p. 329.

Mechanism #coordinate-wise-nominal-neighbor-vote

For each nominal coordinate, assign the value described as occurring in the majority of the k nearest neighbors. This is a vote over feature values within the selected minority-neighbor set, not a vote to predict the class of a query example. The printed NC rule does not expressly add the seed to this vote. It does not completely specify how to choose a category when frequencies tie or when no category has a strict majority.

supports Source: SMOTE: Synthetic Minority Over-sampling Technique · Locator: Section 6.1 step 3, PDF p. 29 / printed p. 349. The separate Section 6.2 proposal explicitly votes over the seed plus its k neighbors, PDF p. 31 / printed p. 351; that voting-set statement is not imported into NC.

⚠️ Curated Limitations & Negative Results

The continuous mechanism follows the shared-scalar prose and Table 1 interpretation. Algorithm SMOTE line 21 draws gap inside the attribute loop; independent coordinate draws generally yield a point in the coordinate-wise rectangle rather than on the line segment. The source descriptions are not certified as identical executable code, and Section 6.1 does not resolve this inherited conflict.
The nominal rule says majority among the k neighbors. It does not completely specify ties, categories lacking a strict majority, equal distances, duplicate observations, adequate neighbor counts, self-inclusion or neighbor reuse. The SMOTE-N proposal explicitly includes the seed in its vote; that distinct statement is not asserted for NC. Oversampling-percentage conventions also require care: the earlier pseudocode handles percentages below 100 by selecting a subset and otherwise assumes integral multiples of 100.
Without a continuous feature the median scale is undefined, so this rule is not an all-nominal method. If the median standard deviation is zero, all nominal mismatch penalties vanish. Distinct vectors with equal continuous coordinates can then have zero distance, even though some nominal coordinates differ. A zero median does not require every continuous feature to be constant. These are direct formula boundaries, not reported experiments; no fallback is printed.
Raw continuous squared differences and a median across feature-wise standard deviations depend on feature units. Independent rescaling can change neighbor selection; this is not feature-wise standardization or arbitrary-unit invariance. The method description does not fix sample versus population standard-deviation denominators or missing and nonfinite-value handling.
Continuous interpolation and coordinate-wise nominal votes do not guarantee a distinct synthetic record, a plausible joint feature combination or a point in the true minority support. The minority label is assigned by the synthesis model rather than newly observed. Nominal votes may select categories from neighbors other than the one selected for the continuous interpolation.
Section 6.1 reports one mixed-feature dataset, Adult, and describes unfavorable comparison with plain undersampling. Its explanation involving overlap with majority space is a hypothesis, not a demonstrated general cause. No numerical performance comparison, plot quantification, benchmark reproduction, universal benefit or calibration guarantee is asserted here.
This record is scoped to the original 2002 construction. Later one-hot encodings, category penalties, voting sets and zero-scale or missing-value fallbacks need separate versioned evidence. SMOTE-N remains a separately described potential nominal-only proposal, not an alias. The retained 37-page journal PDF was acquired on 11 October 2026 and reused after byte verification. Research read the defining core and extensions with selected formula visuals, not the complete paper or benchmark. No author code, later library implementation or original VDM paper was inspected or executed. The mixed-distance example is Table 6 despite the preceding prose reference to Table 2.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:DataAugmentation` Data Augmentation	broad_match	SMOTE-NC constructs additional mixed-feature training examples from observed minority examples through continuous interpolation and nominal-neighbor voting. The pinned data-augmentation concept is broader than this particular preprocessing method. This is a local-to-broader mapping, not equivalence, an upstream ontology edit, a neural preprocessing layer or a guarantee of improved robustness or predictive accuracy.

Evidence Sources & Primary Literature

1 Source

Source ID	Publication / Document	Type / Access	Reference & Checksum
`chawla2002`	SMOTE: Synthetic Minority Over-sampling Technique Nitesh V. Chawla, Kevin W. Bowyer, Lawrence O. Hall, W. Philip Kegelmeyer	primary paper full_text	ID: https://doi.org/10.1613/jair.953 sha256:8eece9837dbc…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-15T07:33:19Z

Updated: 2026-10-15T07:33:19Z

Source File: data/curated/SMOTENC.yaml