SMOTE extensions for abnormal minority instances in imbalanced time series classification

Class imbalance remains a significant challenge in time series classification, especially when the minority samples include abnormal or outlier instances. This study proposes four extensions of the Synthetic Minority Oversampling Technique (SMOTE): Distance Extended SMOTE, Dirichlet Extended SMOTE, Feature-Correlation and Random-Perturbation SMOTE, and Bayesian Gaussian Mixture Model (BGMM) SMOTE. These methods use weighted neighbor selection, probabilistic modeling, and clustering to generate synthetic samples. Augmentation was applied only to the training data after train–test splitting and within the validation process, while the test data remained unseen. The proposed methods were integrated with a Bayesian Deep Bidirectional LSTM classifier using Monte Carlo dropout-based variational inference to enable both robust classification and uncertainty quantification. Experiments on five benchmark time-series datasets showed that based on stratified 10-fold cross-validation, the proposed SMOTE extensions consistently surpassed baseline methods and standard SMOTE. BGMM-SMOTE and Dirichlet Extended SMOTE achieved the best performance. The present study gives a detailed analysis of SMOTE enhancements specifically designed for abnormal class distributions, providing an appropriate method to handle imbalanced time series classification. The framework combines advanced resampling methods with Bayesian deep learning to make more reliable predictions, with applications in healthcare monitoring, fault detection, and other safety-critical fields.