ReCoG: Relational and Compact Context Graph Learning for Few-shot Molecular Property Prediction

Few-shot molecular property prediction (FSMPP) is essential in drug discovery and materials design, where high-quality labeled data are often scarce and expensive to obtain. Despite the promising performance of existing methods, especially context-aware methods, they still face two-fold severe challenges with $\mathit{\text{insufficient}}\mathit{\text{structural}}\mathit{\text{context}}\mathit{\text{modeling}}$ & $\mathit{\text{redundant}}\mathit{\text{auxiliary}}\mathit{\text{context}}\mathit{\text{learning}}$, leading to inadequate context graph exploration and ineffective information utilization for effective molecule representation learning. To address these, in this paper, we propose a novel framework by learning on $\underset{‾}{\mathbf{\text{Re}}}$lational and $\underset{‾}{\mathbf{\text{C}}}$ompact c$\underset{‾}{\mathbf{\text{o}}}$ntext $\underset{‾}{\mathbf{\text{G}}}$raph, named $\mathbf{\text{method}}$, to comprehensively exploit the context graph for expressive molecular property prediction. Specifically, the proposed \method contains two core modules: a $\mathbf{\text{(1)}}\mathbf{\text{cross-property}}\mathbf{\text{relational}}\mathbf{\text{learning}}\mathbf{\text{module}}$ to better model the structural and relational context information, and a $\mathbf{\text{(2)}}\mathbf{\text{context}}\mathbf{\text{graph}}\mathbf{\text{information}}\mathbf{\text{bottleneck}}\mathbf{\text{module}}$ to adaptively suppress irrelevant auxiliary signals for compact context information utilization, followed by a detailed theoretical demonstration regarding the importance of joint relational and compact knowledge extraction in context graphs.