Synthetic Oversampling Calculator

Oversampling Results

Configure the dataset and run the calculator to view results.

1. Dataset Input

Paste dataset

CSV, tab, semicolon, or space-delimited data is supported.

Upload CSV

Delimiter

Header row

Target column

Feature columns

Use Ctrl or Command to select multiple columns.

Ignored columns

Categorical features

Priority class

Dataset preparation

2. Oversampling Configuration

Oversampling method

Sampling strategy

Classes to oversample

Target ratio

Target count

Increase percentage

Per-class targets

Nearest neighbors (k)

Borderline neighbors (m)

Distance metric

Minkowski power

Random seed

Interpolation strength

Noise or jitter

ADASYN beta

KMeans clusters

Cluster balance threshold

SVM kernel

SVM C

SVM gamma

3. Preprocessing and Data Quality

Numeric scaling

Missing values

Outlier handling

Categorical encoding

Duplicate rows

Constant features

4. Validation Workflow

Train-test handling

Training percentage

Cross-validation folds

CV repeats

Cross-validation mode

Oversample inside each fold

Oversampling before splitting can leak synthetic information into validation data. Training-only resampling is recommended.

5. Optional Model Impact Metrics

Enter confusion-matrix counts to compare evaluation metrics. These values are independent from synthetic sample generation.

Before TP

Before FP

Before TN

Before FN

After TP

After FP

After TN

After FN

6. Run Calculator

Exports and Generated Code

Implementation language

Run the calculator to generate reproducible implementation code.

Formula Used

SMOTE creates a sample between a minority observation and one neighbor. It uses a random interpolation coefficient. The new point remains within their feature segment.

x_new = x_i + λ × (x_neighbor − x_i), where 0 ≤ λ ≤ 1

ADASYN allocates more samples to difficult minority regions. Borderline methods emphasize observations near competing classes. Random oversampling duplicates existing minority observations with optional jitter.

How to Use

Paste or upload a labeled dataset.
Parse it and confirm target, feature, and categorical columns.
Select an oversampling method and target strategy.
Choose preprocessing, validation, and algorithm parameters.
Generate samples, inspect warnings, and export results.

Example Data

age	income	tenure	region	label
25	42000	1	North	No
30	54000	3	East	Yes

Frequently Asked Questions

What is synthetic oversampling?

It creates new minority observations instead of only duplicating rows. Generated values follow patterns between nearby minority samples. This can improve model learning on rare classes.

When should SMOTE be used?

Use SMOTE for imbalanced classification with meaningful numeric neighborhoods. Apply it only within training data. Validate results using class-sensitive evaluation metrics.

What does SMOTE-NC handle?

SMOTE-NC supports mixtures of numerical and categorical features. Numeric values are interpolated between neighbors. Categories are selected from nearby minority observations.

Why can oversampling cause leakage?

Leakage occurs when synthetic data uses validation observations. Split data before creating synthetic samples. Pipeline-based resampling prevents this common evaluation error.

How should neighbor count be chosen?

Small minority classes require a low neighbor count. Larger values smooth the generated distribution. Cross-validation helps compare reasonable settings safely.

What is ADASYN best for?

ADASYN emphasizes minority points surrounded by competing classes. It focuses generation on difficult regions. This can also amplify noisy observations.

Which metrics suit imbalanced data?

Prefer recall, F1, balanced accuracy, G-mean, and PR-AUC. Accuracy alone can hide minority failures. Compare several metrics across validation folds.

Can multiclass data be oversampled?

Yes, each eligible class can receive a separate target. The calculator supports per-class mappings. Review class overlap after generating samples.

Should test data be balanced?

Usually test data should preserve its natural distribution. Oversampling is intended for model training. Realistic testing gives more trustworthy performance estimates.