Imbalanced classification poses a persistent challenge in real-world decision-support systems, where the rare positive class is typically the focus of interest and predicted probabilities directly influence critical decisions. Many conventional oversampling methods rely on uniform interpolation between minority instances, often generating redundant or noisy samples that fail to capture the heterogeneous structure and local uncertainties inherent in real minority classes. This study introduces LSEA (Latent-Structure Entropy Allocation), a reliability-oriented approach that models the minority class using a Gaussian Mixture Model and employs hierarchical entropy, at both component and instance levels, to direct synthetic sample generation toward structurally ambiguous regions.Evaluated on ten KEEL benchmark datasets and two real-world applications (credit card fraud detection and breast cancer diagnosis), LSEA delivers precision-recall performance and probability calibration comparable to established baselines, including SMOTE-family methods, GMF-SMOTE, and DeepSMOTE. Statistical tests show no significant pairwise differences among the seven methods in most cases. Unlike deep generative alternatives, LSEA achieves these results without neural network training, GPU resources, or extensive hyperparameter tuning, while offering greater interpretability through its explicit entropy-guided allocation.These findings suggest that combining latent structure modeling with entropy-based prioritization provides a practical and reliable way to improve learning under class imbalance, particularly when calibration and computational efficiency matter.