Meta–Graph Prototypical Diffusion for Tabular Classification

Boshko Koloski, Nada Lavrač, Blaž Škrlj

Submission Type: Short paper (4 pages)

Keywords: tabular classification; graph learning; few-shot learning; high-dimensional classification; meta learning

Abstract: High-dimensional, low-sample-size (HDLSS) classification is a persistent challenge in microarray genomics and related domains, where limited samples, noisy features, and unreliable high-dimensional geometry confound standard methods. We propose MGPD, a meta-learning approach that learns compact embeddings, constructs and fuses multiple complementary graph representations with learned weights, and combines prototype-based global reasoning with local neighbor evidence via APPNP diffusion on strictly inductive soft-kNN graphs. Our hybrid readout incorporates balanced priors to handle class imbalance. On six microarray benchmarks, MGPD attains balanced accuracy on par with state-of-the-art general-purpose methods (RealMLP, TabPFN v2, TabICL) while achieving superior average AUPRC, demonstrating that compact, inductive graph-based architectures can compete with heavily pretrained tabular models on HDLSS tasks.

Submission Number: 55