SiameseNet+: Classification of HDLSS Data with Tensor Affinities

Building reliable classification models for high-dimensional, low-sample size (HDLSS) data is challenging due to the prevalence of the concentration effect and the risk of overfitting. These issues are especially pronounced when using conventional Siamese networks that primarily rely on pairwise affinities. To address these limitations, a novel framework named SiameseNet+ is proposed for HDLSS classification. Specifically, this framework incorporates high-order tensor affinities, extending beyond traditional pairwise interactions, into a Siamese network architecture for capturing richer spatial interactions among HDLSS samples. By jointly modeling pairwise, ternary, and quaternary relations through tensor operations, the framework enables geometry-aware and stable representation learning while mitigating overfitting in HDLSS regimes. To ensure scalability, a computationally efficient mini-batch training strategy is adopted to reduce computational overhead while preserving sufficient multi-sample information coverage. Extensive experiments on twelve real-world HDLSS benchmarks and high-dimensional simulations demonstrate that the proposed SiameseNet+ achieves competitive and robust performance compared with twelve representative baselines.