Robust Chemical Reaction Condition Recommendations via Label Mix Strategy

Recommending optimal reaction conditions remains a core challenge in AI-driven chemistry due to the limited representation of condition features and the sparsity of labeled data. We propose a collaborative filtering framework that encodes reaction conditions using graph neural networks and captures their interaction with reactions through a differential cross-attention mechanism, which selectively focuses on the most relevant condition features. To mitigate noise and sparsity in the data sets, we introduce a Label Mix strategy that augments training signals by blending condition labels, along with co-occurrence matrix regularization to encourage chemically meaningful interactions. Extensive evaluations on multiple benchmark data sets demonstrate the effectiveness of our approach in both full-data and few-shot regimes, with consistent improvements in accuracy, generalization, and robustness.