Data Augmentation for Tabular Data

A simple augmentation technique where values for a specific feature are randomly swapped between different records in the dataset. This preserves the global marginal distribution of the feature while breaking specific record-level correlations.

• Preserves Global Statistics: The overall frequency of each value remains unchanged.
• Introduces Local Variation: Creates new, plausible combinations of feature values that may not exist in the original dataset.
• Use Case: Effective for increasing diversity in datasets for models like decision trees or gradient boosting machines that learn rules based on feature splits.

The addition of small, random perturbations to continuous numerical features. This technique helps models become robust to minor variations and measurement errors present in real-world data.

• Gaussian Noise: Adds values drawn from a normal distribution N(0, σ), where σ is a small standard deviation relative to the feature's scale.
• Uniform Noise: Adds values from a uniform distribution within a defined range [-a, a].
• Impact: Regularizes the model, preventing it from overfitting to exact numerical values and improving generalization to unseen data.

A privacy-enhancing and robustness technique where the order of non-ordinal columns (features) in the dataset is randomly shuffled. This breaks any spurious correlations that might be learned from column position rather than actual semantic meaning.

• Targets Model Assumptions: Challenges models that might inadvertently rely on the positional index of a feature.
• Privacy Benefit: Can help obscure the identity of sensitive features in a dataset schema.
• Limitation: Not applicable when feature order is meaningful (e.g., time series data, sequence data).

A specialized augmentation algorithm designed to address class imbalance. It generates synthetic examples for the minority class by interpolating between existing, similar instances in feature space.

• Process: For a minority class instance, SMOTE finds its k-nearest neighbors, selects one randomly, and creates a new synthetic point along the line segment connecting them.
• Key Benefit: Mitigates overfitting caused by simple duplication (oversampling) by creating new, plausible data points.
• Consideration: Primarily designed for continuous features; variants like SMOTE-NC exist for handling categorical data.

The generation of new records by applying domain-specific business rules or logical constraints to existing data. This ensures synthetic data adheres to real-world plausibility checks.

• Examples: Ensuring a total_price column equals unit_price * quantity, or that an employment_start_date is before an employment_end_date.
• Advantage: Produces highly realistic data that respects known data invariants and integrity constraints.
• Implementation: Often uses if-then logic, probabilistic graphical models, or small decision trees defined by a domain expert.

Randomly changing the category of a discrete feature to another valid category based on a defined probability or distribution. This increases diversity in categorical feature representations.

• Uniform Flipping: Changes a category to any other with equal probability.
• Distribution-Aware Flipping: Changes a category based on the empirical marginal distribution of that feature (more common categories are more likely targets).
• Application: Useful for simulating entry errors or natural variation in categorical responses (e.g., a user occasionally selecting a different product category).

A primary use case is addressing class imbalance in classification tasks. When one class is underrepresented, models become biased. Simple augmentation techniques like SMOTE (Synthetic Minority Over-sampling Technique) generate synthetic examples for the minority class by interpolating between existing neighbors in feature space. This creates a more balanced training set without collecting costly real-world data, leading to models with better recall for rare but critical events like fraud detection or equipment failure.

Augmentation acts as a regularizer, preventing overfitting by exposing the model to more varied data. Techniques include:

• Adding controlled noise (e.g., Gaussian) to continuous features.
• Swapping or permuting values within a column (for non-ordinal categories).
• Creating missing values at random and having the model learn to handle them. This teaches the model to focus on robust patterns rather than memorizing specific training examples, improving performance on unseen, real-world test data where distributions may shift slightly.

Rule-based augmentation can create de-identified datasets for safe sharing. By applying transformations like:

• Value masking or binning for sensitive columns (e.g., age ranges instead of exact age).
• Data shuffling across records to break direct identifiers.
• Adding differential privacy noise to aggregated statistics. This generates a useful, statistically similar dataset for development or testing without exposing personally identifiable information (PII), facilitating collaboration and compliance with regulations like GDPR.

Engineers use tabular augmentation to generate realistic test data for application and pipeline development. By applying business logic rules (e.g., 'order total = quantity * price') and statistical distributions learned from production schemas, they can create vast volumes of synthetic test fixtures. This enables:

• Stress testing database performance.
• Validating ETL (Extract, Transform, Load) logic.
• Developing front-end applications without access to live production data, accelerating development cycles while maintaining security.

In domains where data collection is expensive or slow (e.g., clinical trials, manufacturing defects), the initial dataset may be too small for effective model training. Augmentation techniques like bootstrapping (sampling with replacement) and creating perturbed copies of existing records can artificially expand the dataset. This provides more examples for the model to learn from, improving convergence and stability during training, especially for complex models like gradient-boosted trees or deep neural networks that require sufficient data to generalize.

Augmentation can be used strategically to stress-test models by generating rare or extreme scenarios. For example:

• Creating records where feature combinations are at the boundaries of observed ranges.
• Simulating data entry errors (e.g., typos in categorical fields, outliers in numerical fields).
• Generating examples that sit directly on the decision boundary between classes. Evaluating model performance on these augmented edge cases helps identify failure modes and improve model resilience before deployment into production environments.

A Generative Adversarial Network specifically architected for tabular data with mixed data types (continuous and categorical). It introduces mode-specific normalization for numerical columns and uses a conditional training strategy via a training-by-sampling technique to effectively model imbalanced categorical distributions. This allows it to generate realistic synthetic rows that preserve complex column-wise relationships.

A Variational Autoencoder adapted for structured data. TVAE uses a Gaussian Mixture Model (GMM) as the prior distribution in its latent space, which is better suited for capturing the multi-modal distributions common in tabular data. The reconstruction loss is specifically designed to handle mixed data types, making it a robust probabilistic alternative to GAN-based tabular generators.

A classical oversampling algorithm used to address class imbalance. For each minority class instance, SMOTE generates synthetic examples by:

• Identifying its k-nearest neighbors within the same class.
• Randomly selecting one neighbor.
• Creating a new point along the line segment in feature space connecting the original instance and the neighbor. It's a foundational, non-deep learning approach for data augmentation in classification tasks.

A rigorous mathematical framework that provides a privacy guarantee for synthetic data generation. An algorithm is differentially private if the inclusion or exclusion of any single individual's record in the input dataset has a negligible statistical effect on the output synthetic data. Methods like PrivBayes inject calibrated noise into the parameters of a learned Bayesian network before sampling, ensuring the synthetic data cannot be used to infer information about any specific person in the original dataset.

The primary evaluation protocol for assessing the utility of synthetic tabular data. The core procedure is:

1. Train a downstream machine learning model (e.g., a classifier) exclusively on the synthetic dataset.
2. Evaluate the trained model's performance on a held-out real dataset. If the model performs nearly as well as one trained on real data, it indicates the synthetic data has high fidelity and preserves the statistical patterns necessary for the task.

A metric for comparing probability distributions, crucial for evaluating synthetic tabular data fidelity. Also called the Earth Mover's Distance, it measures the minimum "cost" of transforming one distribution into another. In practice, the 1-Dimensional Wasserstein Distance is often calculated per feature to assess how well the synthetic data's marginal distributions (for both continuous and categorical columns) match the real data's. Lower distances indicate higher distributional similarity.