This study proposes a hybrid generative–ensemble framework to predict key mechanical properties of recycled aggregate concrete from mix proportions. An established database of 112 mixes was used to model compressive strength, split tensile strength, flexural strength, and elastic modulus. To mitigate data scarcity, a conditional variational autoencoder was trained on the training data only and used to generate additional physically plausible input samples, after which seven supervised learning algorithms were trained and compared using cross-validation. Gradient boosting and support vector regression achieved the most accurate and stable predictions across all targets, outperforming baseline linear models and commonly used empirical correlations. Feature-attribution analysis was used to identify the dominant drivers of each property, showing that binder-related variables primarily govern strength, while aggregate-related variables dominate stiffness. The results support practical, data-driven screening of recycled aggregate concrete mixes and provide interpretable guidance for sustainable mix design.