INDEQS: Informed Neural controlled Differential EQuationS
INDEQS: Informed Neural controlled Differential EQuationS
Authors: Michael Detzel, Gabriel Nobis, Kristiyan Blagov, Juri Schubert, J Y, Wojciech Samek Date: 2026-06-17 Paper ID: openalex:2606.19138
Summary
INDEQS is a graph-based Neural Controlled Differential Equation (NCDE) framework designed to integrate prior knowledge of directed graph structures into continuous-time time series forecasting. The method separates inner state mixing across graph nodes from outer mixing between the vector field and control, allowing for more structured and parameter-efficient learning. Evaluated on synthetic advection simulations and real-world river discharge and traffic flow (PeMS08) data, INDEQS consistently outperforms standard uninformed NCDEs, particularly when graph informedness is applied to the outer mixing components.
Key Contributions
- Introduces INDEQS, a framework that incorporates known directed graph structures into NCDEs by separating inner and outer state mixing.
- Proposes two variants: a lightweight graph-constrained approach and an expressive variant with adaptive graph convolutions.
- Demonstrates superior forecasting performance on river discharge networks and the PeMS08 traffic dataset compared to uninformed NCDE baselines.
Key Concepts
- indeqs-framework: A graph-based Neural Controlled Differential Equation framework that injects prior directed graph knowledge through the architectural separation of inner state mixing and outer control-vector field mixing.
Archivist Review
The paper introduces a structured way to incorporate prior knowledge into neural controlled differential equations. I have approved the framework as it proposes a specific, reusable architectural design (separating inner/outer mixing) for continuous-time spatio-temporal modeling. PeMS08 is approved as a representative dataset for traffic flow prediction.
Approved Concepts
- Informed Neural Controlled Differential Equations (INDEQS): It defines a systematic methodology for injecting known relational structure into neural differential equations by bifurcating the integration of graph connectivity from the vector field-control interaction.
Rejected Candidates
- [concept] INDEQS (
indeqs) - other: Renamed for better descriptive clarity and consistency with established naming patterns in the vault.
Datasets
Links
Metadata & Links
- url
- https://arxiv.org/abs/2606.19138
- paper_id
- 2606.19138
- paper_source
- openalex
- domain
- time-series
- tags
- time-series-forecastinggraph-neural-networksneural-differential-equationsspatio-temporal-forecastingphysics-informed-learning
- architectures
-
- datasets
- pems08
- concept_slugs
- indeqs-framework
- dataset_slugs
- pems08
- skill
- TimeSeriesSkill
- created_at
- 2026-06-20T06:32:56Z