Research map / Probabilistic and scientific modeling / Dynamical systems

Continuous Temporal Domain Generalization

NeurIPS 2024 · Probabilistic and scientific modeling › Dynamical systems · approach: Latent dynamics models · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces continuous temporal domain generalization (CTDG) and the Koodos framework, which models model-parameter dynamics with a Koopman operator inside a neural ODE-style system to generalize to arbitrary future times.

Problem

Temporal domain generalization assumes domains at fixed discrete time steps, but real data arrive at irregular continuous times; the task is to predict model parameters at arbitrary unseen times under gradual concept drift.

Approach

Treat predictive-model parameters as evolving under a continuous dynamical system, map them into a learned latent space via an autoencoder where dynamics are linear (Koopman operator), and jointly optimize prediction, reconstruction, dynamic fidelity and consistency losses; spectral constraints on the Koopman operator enable stability analysis and prior-knowledge control (e.g., skew-symmetric K fo

Key results

• On continuous benchmarks Koodos reports 2.8% error on 2-Moons versus 12.8% for the best baseline V-REx and 4.6% on Rot-MNIST versus 6.6% for Offline.
• Koodos achieves the best Twitter AUC of 0.71 and the best Cyclone MAE of 16.4 and House MAE of 9.0 among the compared methods.
• On discrete TDG benchmarks Koodos reports 1.3% error on 2-Moons versus 3.2% for DRAIN and MAE 4.8 on Appliance versus 6.4 for DRAIN.

Measured effects reported in the paper

Surprising findings

• Discrete-time baselines such as DRAIN and CIDA performed poorly on the continuous benchmarks, with some worse than the simple Offline model.
• Learned Koopman operators can have eigenvalues with positive real part, so extremely long-horizon extrapolation diverges unless the operator is constrained to be skew-symmetric.

Limitations

• Assumes smooth, predictable concept drift; other drift types need alternative frameworks.
• Long-horizon extrapolation with a learned unconstrained Koopman operator diverges and accumulates errors.
• Theoretical guarantee is informal and rests on strong assumptions (invertible Jacobian, Lipschitz drift).

Experimental setup

Datasets: Rotated 2-Moons Rotated MNIST Twitter Influenza Risk Yearbook Tropical Cyclone Intensity House Prices Online News Popularity Shuttle Appliances energy

Benchmarks: Rotated 2-Moons Rotated MNIST Twitter Yearbook Cyclone House ONP Shuttle Appliance

Metrics: error rate (%) AUC MAE

Baselines: Offline LastDomain IncFinetune IRM V-REx CIDA TKNets DRAIN DRAIN-Δt DeepODE CDOT GI

Builds on: Neural ODE Koopman theory DRAIN DeepODE TKNets

Closest papers by meaning

• Generalizing across Temporal Domains with Koopman Operators (AAAI 2024 · similarity 0.94)
• Learning Robust Spectral Dynamics for Temporal Domain Generalization (NeurIPS 2025 · similarity 0.90)
• Continuous Domain Generalization (NeurIPS 2025 · similarity 0.89)
• Learning Time-Aware Causal Representation for Model Generalization in Evolving Domains (ICML 2025 · similarity 0.87)
• Enhancing Evolving Domain Generalization through Dynamic Latent Representations (AAAI 2024 · similarity 0.86)
• Out-of-Domain Generalization in Dynamical Systems Reconstruction (ICML 2024 · similarity 0.86)