<p>Accurately understanding the mechanical properties of natural gas hydrate-bearing sediment (GHBS) is critical for the safe exploitation of marine natural gas hydrate. However, directly linking the observed physical properties of GHBS to their mechanical behaviour remains a significant challenge. Theoretical hydrate-modified constitutive models are limited by their complexity in parameterisation and incomplete representation of underlying patterns, while data-driven models face the challenge of small-sample datasets from GHBS mechanical experiments. This study integrates the knowledge of a theoretical hydrate-modified constitutive model and experimental data, to develop a data-driven model for representing the mechanical behaviour of GHBS using the transfer learning method. The performance of the proposed model is validated on a test set containing both representative coarse-grained GHBS and fine-grained GHBS sample data. The results indicate that the model can accurately predict the mechanical behaviour of a wide range of GHBS samples under different conditions, including strain hardening/softening and shear contraction/dilatation phenomena. Ablation experiments, conducted by simplifying the model structure or loss function, demonstrate that the transfer learning algorithm employed in this study significantly improves prediction accuracy compared to conventional deep learning networks. The generalisation capability of the proposed model is further validated through comparisons with triaxial test data on hydrate dissociation. Finally, comparisons with existing studies on GHBS mechanical behaviour indicate that the proposed model delivers low cost, broad applicability, and fast, accurate predictions, serving as a practical surrogate for laboratory testing across most GHBS types and laying the groundwork for data-driven constitutive models.</p>