LSD-SMOTE: Latent Semantics Disentangle SMOTE for Imbalanced Data Over-sampling

Class imbalance is a pervasive and challenging issue in machine learning and data mining, where standard classifiers lean heavily toward the majority class. Although the Synthetic Minority Over-sampling Technique (SMOTE) is widely used to mitigate this problem, generating synthetic samples directly in the original high-dimensional feature space often introduces noise, cross-class blending, and unrealistic combinations due to the intricate coupling and non-linear correlations of raw features. To address these limitations, this paper proposes a novel over-sampling framework named Latent Semantics Disentangle SMOTE (LSD-SMOTE).

The proposed method leverages a Variational Autoencoder (VAE) integrated with a beta-disentanglement penalty to map high-dimensional minority samples into a lower-dimensional latent space, forcing the latent variables to decouple into mutually independent semantic factors. By performing linear interpolation within this decoupled latent space, SMOTE can smoothly manipulate individual semantic axes without causing structural distortion. Finally, a trained decoder reconstructs these newly synthesized latent vectors back into high-quality, diverse minority samples within the original feature space. By decoupling the underlying semantics before generation, LSD-SMOTE effectively prevents the generation of ambiguous boundary samples and enhances the classification performance on imbalanced datasets.