Skip to main content
Back to timeline
arXivSource publication:

Integral residuals replace derivatives in higher-order interaction reconstruction: IntegralPINN reaches 0.730 Macro-AUC on social contagion under noisy, sparse observations

Synopsis

Recovering higher-order interactions in networks from noisy, sparse time series has been limited by how differentiation amplifies noise. IntegralPINN replaces the physics residual with an integral over time intervals and parameterizes interaction tensors as bounded learnable variables optimized jointly with a neural trajectory surrogate. Across ecological, social-contagion and neuronal-synchronization systems it attains higher Macro-AUC than HyperPINN and other baselines under noisy observations, reaching 0.730 versus 0.585 on social contagion.

Source-provided article image: Higher-Order Interaction Reconstruction via Integral Physics-Informed Neural Networks
arXiv

Interpretation

Higher-order interaction reconstruction no longer relies on differentiating a noisy trajectory; an integral residual over time intervals enforces consistency with the governing dynamics. Earlier physics-informed networks such as HyperPINN penalized differential residuals that match state derivatives, and derivatives of a surrogate fitted to noisy discrete observations amplify high-frequency error. On ecological population, social contagion and neuronal phase synchronization systems, IntegralPINN's Macro-AUC exceeds HyperPINN, the kernel-based method, THIS and integral SINDy; on social contagion IntegralPINN reaches 0.730 against 0.585 for HyperPINN.

A neural surrogate supplies a continuous state trajectory, while higher-order interaction tensors are bounded learnable parameters optimized jointly with that trajectory by gradient descent. The integral residual swaps derivative matching for state increments and integrated vector fields over an interval, avoiding explicit differentiation and letting binary observations enter through a cross-entropy loss. Experiments cover ecological dynamics with continuous observations, a social contagion process with 300 binary snapshots, and a 30-node multilayer Kuramoto network whose slow phase carries simplicial interactions up to order five.

Perspective

Validation is limited to small- and medium-scale simulated systems: ecological populations, social contagion and a 30-node neuronal synchronization network, with node dynamics and higher-order interaction function forms assumed known.

The theoretical error bound holds under assumptions that the surrogate fits accurately and the integration window satisfies an excitation condition, and the analysis fixes the surrogate and window length without accounting for bounded parameterization, sparsity regularization or joint optimization; excessively long windows weaken the excitation condition, so the integral form does not uniformly outperform the differential one.