Synthetic Minority Over-sampling Technique (SMOTE)

SMOTE synthesizes new minority class samples by operating in the feature space rather than the raw signal space. For each minority sample, the algorithm identifies its k-nearest neighbors and randomly selects one. A new synthetic vector is created at a random point along the line segment connecting the original sample and its chosen neighbor. This forces the decision region of the classifier to become larger and less specific, combating overfitting.

• Key Parameter: k_neighbors controls the granularity of interpolation.
• Operation: Works on extracted features (e.g., cumulants, spectral coefficients), not raw IQ samples directly.
• Benefit: Prevents a classifier from memorizing sparse, isolated minority examples.

A naive approach to class imbalance is random oversampling, which duplicates existing minority samples verbatim. This leads to severe overfitting, as the model memorizes exact copies rather than learning generalizable patterns. SMOTE mitigates this by generating non-replicated, synthetic variations that share statistical similarities with the parent class but are not identical.

• Problem: Exact duplication creates sharp, brittle decision boundaries.
• SMOTE Solution: Synthetic interpolation broadens the minority class distribution.
• Result: Improved generalization to unseen test data and real-world RF channel conditions.

In RF machine learning, certain modulation types or emitter signatures are inherently rare in the wild. SMOTE is applied to the feature vector extracted from IQ samples to balance the training set before feeding it to a classifier.

• Example: Balancing a dataset where BPSK signals appear 10,000 times but QAM64 appears only 50 times.
• Features Used: High-order cumulants, spectral symmetry measures, and wavelet coefficients.
• Caveat: Applying SMOTE directly to raw time-series IQ data can generate physically invalid signals; feature-level application is standard.

Standard SMOTE blindly synthesizes samples from all minority instances, including those deep within the cluster core. Borderline-SMOTE improves this by only oversampling minority examples that lie near the decision boundary—those misclassified or adjacent to majority class samples.

• Mechanism: Classifies minority samples as 'safe', 'danger', or 'noise' based on neighbor composition.
• Focus: Synthesizes only from 'danger' samples to sharpen the boundary.
• RF Use Case: Distinguishing between spectrally similar modulations like 16-QAM and 64-QAM where the boundary is ambiguous.

While SMOTE is a foundational geometric technique, it is often compared and combined with deep generative models like Variational Autoencoders (VAEs) and GANs for RF data augmentation. SMOTE provides a lightweight, deterministic baseline, whereas GANs learn the full probability distribution of the signal.

• SMOTE: Fast, deterministic, operates on existing feature vectors.
• GAN/VAE: Stochastic, learns to generate entirely new feature vectors from latent space.
• Hybrid Approach: Use SMOTE for quick class rebalancing during exploratory analysis; deploy GANs for high-fidelity synthetic signal generation in production.

SMOTE's effectiveness degrades in very high-dimensional spaces due to the curse of dimensionality. As the number of features increases, the concept of a 'nearest neighbor' becomes less meaningful, and linear interpolation may generate samples that do not lie on the true data manifold.

• Mitigation: Apply dimensionality reduction (e.g., PCA or t-SNE) before SMOTE.
• Alternative: Use manifold-aware generative models like Diffusion Models for raw high-dimensional data.
• Noise Amplification: SMOTE can amplify existing noise if outliers are not cleaned beforehand.