<p>Polymer informatics faces challenges owing to data scarcity arising from complex chemistries, experimental limitations, and processing-dependent properties. This review presents the recent advances in data-efficient machine learning for polymers. First, data preparation techniques such as data augmentation and rational representation help expand the dataset size and develop useful features for learning. Second, modeling approaches, including classical algorithms and physics-informed methods, enhance the model robustness and reliability under limited data conditions. Third, learning strategies, such as transfer learning and active learning, aim to improve generalization and guide efficient data acquisition. This review concludes by outlining future opportunities in machine learning for small-data scenarios in polymers. This review is expected to serve as a useful tool for newcomers and offer deeper insights for experienced researchers in the field.</p>