<p>Combinatorial pattern learning plays an increasingly crucial role across various domains. Nonetheless, the inherent discreteness and complex structures of combinatorial data, together with the contextual interdependencies intrinsic to these tasks, pose fundamental challenges that necessitate specialised architectures to enable effective training and robust inference. In this study, we present a generalised dual-graph broad feature inference model (GDBIM) to address combinatorial data inference challenges. The dual-graph architecture effectively captures both the inherent interplays among inputs and the potential dependencies among prediction targets in multiple channels. Furthermore, by employing a feature expansion strategy, the model further extracts complex feature patterns, thereby facilitating targeted inference. Experiments across diverse disciplines demonstrate that GDBIM achieves robust performance in both combinatorial regression and classification tasks. Both ablation and parameter analyses have been conducted to substantiate the design rationales behind GDBIM further.</p>