The advent of machine learning has revolutionized building structure-property relationships and materials discovery. However, applying these tools to polymers is challenging due to complex interactions, vast combinatorial spaces, and multiscale behavior. We propose HAPPY (Hierarchically Abstracted rePeat unit of PolYmers), a string-based representation that groups substructures with grammatically complete connectors. Combined with RNN or Transformer models, HAPPY achieves accurate property prediction even with limited data and supports inverse design of target-property polymers, while offering chemical interpretability that links specific substructures to target properties. We also developed a complementary graph-based model to capture chain architectures and complex connectivity. Integrating hierarchical abstraction, deep sequence modeling, and graph analysis, we present a versatile, data-efficient platform for next-generation polymer informatics, with principles that extend naturally to broader classes of structurally complex molecular systems.