SphUnc: Hyperspherical Uncertainty Decomposition and Causal Identification via Information Geometry

SphUnc, a novel machine learning framework, integrates hyperspherical representation learning with structural causal modeling to enhance uncertainty decomposition and causal identification within multi-agent systems. By utilizing von Mises-Fisher distributions, the model translates features into unit hypersphere latents, separating uncertainty into epistemic and aleatoric elements through information-geometric fusion. This framework employs a structural causal model on spherical latents, facilitating the identification of directed influences and enabling interventional reasoning via sample-based simulations. Empirical tests on social and affective benchmarks demonstrate improved accuracy, superior calibration, and interpretable causal signals, laying a geometric-causal groundwork for uncertainty-aware reasoning amid higher-order interactions in multi-agent environments. The paper was submitted to arXiv on March 3, 2025.