Tabular data learning improvement via behavior molecule structure denoising

The reliance of Graph Neural Networks on reliable graph structures limits their performance when confronted with noisy, incomplete, or erroneous data. To tackle this challenge, we introduce a novel method that models behavioral tabular data as behavior molecule graphs, capturing fine-grained structural information while mitigating noise propagation. Building on this foundation, we introduce the Behavior Molecule Structure Denoising method within the framework of Graph Structure Learning. Utilizing self-supervised contrastive learning, the method optimizes graph quality through maximizing mutual information between the learned structure and anchor graphs, uncovering latent connections, and reducing noise. Experimental evaluations on real-world datasets demonstrate significant improvements in behavior recognition tasks, underscoring the potential of this unsupervised framework for robust structural optimization and fine-grained graph data modeling.