I am preparing to post a preprint to the cs.LG (Machine Learning) section of arXiv and would be grateful if you would consider endorsing my submission. The article has been peer-reviewed and published:
Mondal, S. HyperSpectrum geometry: a Riemannian hyperspherical framework for curvature-adaptive text classification. Neural Computing and Applications 38, 710 (2026).
Abstract
Most neural text classification models embed text in high-dimensional vector spaces and measure semantic similarity under an implicit assumption of flat (Euclidean) geometry. However, real-world semantic structures, especially in emotionally entangled datasets, exhibit anisotropic correlations and non-uniform separability that flat representations cannot adequately capture.
We propose HyperSpectrum Geometry, a curvature-controlled hyperspherical manifold learning framework that combines a learnable global Riemannian metric tensor with normalized angular hyperspherical classification. Our model maps sentence embeddings into a learned hyperspace, performs a Mahalanobis-style metric deformation, and finally executes angular decision partitioning on a deformed Riemannian hypersphere.
We further study empirical evidence from structurally different datasets, including GoEmotions and AG News. Our results show that semantically entangled emotion datasets yield a much higher learned curvature, whereas structurally separable topic datasets rely primarily on angular partitioning with relatively small metric deformations. This reveals a curvature–separability trade-off: emotion classification benefits from anisotropic manifold deformation, whereas topic classification relies mostly on angular separation.
Across multiple random seeds, the model achieves competitive results relative to baseline methods while producing valuable geometric interpretations such as curvature magnitude, metric condition number, and intra/inter-class angular similarity. Our findings suggest that learnable geometric curvature is a critical and underexplored component of semantic representation learning.
arXiv has asked me to forward the endorsement request to someone already registered as an endorser for cs.LG. If you are comfortable supporting the submission, you can use the link below:
Alternatively, visit Log in to arXiv | arXiv e-print repository and enter the endorsement code: 4AQNQW
Thank you very much for considering this.
Kind regards,
Sayak Mondal
About the journal-
Neural Computing and Applications (Springer Nature, est. 1993) publishes peer-reviewed research on applied neural computing and machine learning (Q1 SJR in Artificial Intelligence and Software, H-index 166).