💻 computer science

INDEQS: Informed Neural controlled Differential EQuationS

INDEQS is a novel graph-based Neural Controlled Differential Equation framework that integrates prior knowledge of directed graphs into distinct architectural components to enhance time series forecasting accuracy and parameter efficiency across synthetic and real-world spatio-temporal tasks.

Original authors: Michael Detzel, Gabriel Nobis, Kristiyan Blagov, Juri Schubert, Jackie Ma, Wojciech Samek

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Michael Detzel, Gabriel Nobis, Kristiyan Blagov, Juri Schubert, Jackie Ma, Wojciech Samek

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the future flow of water in a river system or the movement of cars on a highway. You have a map (a graph) showing how the rivers connect or how the roads link up.

For a long time, computer models trying to predict these flows have been like students who refuse to look at the map. Even though the map exists and clearly shows that "River A flows into River B," these models try to figure out the connections entirely by guessing based on past data. They are smart, but they are ignoring a huge clue.

This paper introduces a new method called INDEQS (Informed Neural controlled Differential EQuationS). Think of INDEQS as a student who finally decides to open the map and use it to make better predictions.

Here is how it works, broken down into simple concepts:

1. The Problem: "Guessing" vs. "Knowing"

Standard AI models for time-series data (like river levels or traffic) usually treat time as a series of snapshots. They try to learn how one point affects another just by looking at the numbers.

The Old Way: If you have a river network, the model has to "learn" that water flows from the top of the mountain to the bottom by analyzing thousands of days of data. It might get it right, but it's inefficient.
The INDEQS Way: This model is "informed." It is handed the map (the graph structure) beforehand. It knows exactly which river feeds into which, or which road connects to which, before it even starts learning.

2. The Two Ways to Use the Map

The paper proposes two different ways to feed this map information into the AI, which they call Inner and Outer informedness.

Outer Informedness (The "Traffic Director"):
Imagine a traffic controller at a busy intersection. They know the layout of the roads (the map). When a car (data) arrives, the controller uses the map to decide how to distribute that car to the next streets.
- How it works: The model uses the known map to guide how information spreads from one node to another after it has processed the data. It allows the model to say, "Okay, I see this data, and because I know the map, I will send this information to these specific neighbors."
- The Result: This is the most flexible and accurate version. It uses the map as a strong guide but still allows the model to learn new patterns if the data suggests something unexpected.
Inner Informedness (The "Strict Rule-Follower"):
Imagine a strict librarian who only allows books to be moved between shelves if there is a direct, pre-approved path.
- How it works: The model is forced to only look at neighbors that are directly connected on the map. It cannot "hallucinate" a connection between two points that aren't linked on the map.
- The Result: This is a lighter, faster, and more efficient model. It is great when you are 100% sure the map is correct and you don't want the AI to make up fake connections. However, it's a bit less flexible if the real world is messy.

3. The "Advection" Simulation (The Training Ground)

To prove their idea works, the authors built a virtual world. They simulated water flowing (advection) along the edges of a graph.

They created "pulses" of water that traveled along edges, split at junctions, and merged at others.
They tested their models on graphs of different sizes (from 4 nodes to 128 nodes).
The Finding: The "Outer" informed model (the Traffic Director) consistently beat the models that didn't use the map, especially as the maps got bigger and more complex. The "Strict Rule-Follower" (Inner) was also good but struggled a bit more when the paths got very long and complex.

4. Real-World Tests

The authors didn't just stop at simulations. They tested INDEQS on two real-world problems:

River Discharge (The Weser River in Germany):
They tried to predict how much water would flow downstream.
- Result: Because rivers naturally follow a strict path (water flows downhill and follows the banks), the "Informed" models worked beautifully. The model that used the map (Outer) predicted the flow better than the model that tried to guess the connections.
- Bonus: They found that using a "continuous" decoder (a way of reading the model's output that treats time as a smooth flow rather than a series of steps) made the predictions even more accurate.
Traffic Flow (PeMS08 Dataset in California):
They tried to predict traffic jams on highways.
- Result: Traffic is messier than rivers. People make random decisions, and accidents happen. The map (road layout) is useful, but it doesn't tell the whole story.
- Finding: Even here, the "Informed" model did slightly better than the "Uninformed" one. It showed that even in chaotic systems, knowing the physical layout helps the AI make better guesses.

5. The Big Takeaway

The paper concludes that knowing the structure helps.

If you have a reliable map of how things connect (like a river network or a power grid), you should tell your AI about it.
The "Outer" approach is the best all-rounder: it uses the map to guide the AI but keeps it flexible enough to learn from the data.
The "Inner" approach is great if you want a smaller, faster model and you are certain the map is perfect.
Using continuous-time models (which treat time as a smooth river rather than a ticking clock) generally leads to better accuracy, especially for real-world tasks.

In short, INDEQS is a smarter way to predict the future of moving things by making sure the AI respects the physical rules of the road (or river) it's traveling on.

Technical Summary: Informed Neural Controlled Differential EQuationS (INDEQS)

Problem Statement

Neural Controlled Differential Equations (NCDEs) offer a powerful continuous-time framework for time series forecasting. However, standard graph-based extensions, such as Spatio-Temporal Graph NCDEs (STG-NCDEs), typically learn spatial structures purely from data. This approach is suboptimal in physical systems (e.g., river networks, road networks) where a directed graph structure is known a priori and encodes causal relationships between spatial positions. Existing methods often fail to leverage this prior knowledge effectively, treating the graph structure as a learnable parameter rather than a fixed constraint or informative prior, potentially leading to inefficient learning or physically implausible predictions.

Methodology

The authors propose INDEQS (Informed Neural Controlled Differential EQuationS), a graph-based NCDE forecasting method that injects prior knowledge of a directed graph into the vector field at two distinct architectural positions: inner and outer.

Architectural Variants

The method distinguishes between two ways graph information enters the dynamics:

Outer Informedness ( $INDEQS_{outer}$ ): The prior graph structure (via a transposed vertex transition matrix $A_{outer}$ ) is applied at the "outer" position, constraining how information from the control path (data) is distributed across nodes before being processed by the hidden state update function. This variant allows for data-adaptive learning of additional connections via adaptive graph convolutions while respecting the base topology.
Inner Informedness ( $INDEQS_{inner}$ ): The graph structure is applied at the "inner" position within the hidden state update function. Here, the adjacency matrix ( $A_{inner}$ ) is hardcoded, strictly constraining the flow of information to only those nodes connected in the known graph. This variant is more parameter-efficient as it removes the learnable node embeddings used in standard adaptive convolutions.

Both variants utilize a coupled system where a hidden state $H(t)$ evolves via an NCDE driven by interpolated data, and a second state $Z(t)$ evolves via a NODE driven by $H(t)$ to capture spatial dependencies. The authors also explore higher-order adjacencies ( $A_V^{(k)}$ ) to incorporate information from predecessors of predecessors.

Simulation Framework

To systematically evaluate when graph informedness is beneficial, the authors developed a continuous advection simulation on directed graphs. Unlike discrete vertex updates, this simulation models quantities moving continuously along graph edges, splitting and merging at nodes according to predefined ratios. This generates synthetic spatio-temporal datasets with known ground-truth flow structures, allowing for controlled experiments on graph size, spatial resolution, and the necessity of prior knowledge.

Decoders

The paper compares discrete convolutional decoders against continuous-time decoders (based on NODEs) for generating future predictions, allowing for flexible prediction at arbitrary time points.

Key Contributions

INDEQS Architecture: Introduction of a graph-informed NCDE that incorporates prior directed-graph information at two specific architectural positions (inner and outer), offering both a parameter-efficient strict-constraint variant and a more expressive data-adaptive variant.
Advection Simulation: Proposal of a continuous-time advection simulation on directed graph edges, yielding synthetic datasets with known ground-truth flow structures to study the benefits of graph informedness in a controlled setting.
Empirical Validation: Comprehensive evaluation across synthetic advection data, real-world river discharge forecasting (hydrological network), and traffic flow prediction (PeMS08).

Results

The evaluation was conducted on synthetic advection data, river discharge data (Weser River), and traffic flow data (PeMS08).

Synthetic Advection: $INDEQS_{outer}$ consistently achieved lower Mean Absolute Error (MAE) than the uninformed STG-NCDE baseline across various graph sizes and spatial resolutions. The performance gap widened with larger graphs. $INDEQS_{inner}$ showed higher MAE on small graphs but outperformed the baseline on larger graphs, suggesting that as graph complexity increases, the ability to learn structure from limited data diminishes, making prior knowledge crucial.
River Discharge: $INDEQS_{outer}$ with a continuous NODE decoder achieved the best performance (MAE 5.01), outperforming both the uninformed STG-NCDE and other informed baselines (DCRNN, GCDE-GRU). Increasing the adjacency matrix power (incorporating upstream information) further improved performance. $INDEQS_{inner}$ offered a parameter-efficient alternative with slightly higher error but faster training times.
Traffic Flow (PeMS08): $INDEQS_{outer}$ slightly outperformed the uninformed STG-NCDE (MAE 16.50 vs. 16.55), demonstrating that graph topology remains useful even in domains with significant stochasticity and human behavioral noise.
Decoders: Continuous-time decoders consistently yielded better accuracy and temporal flexibility on real-world tasks compared to discrete convolutional decoders, though they incurred higher computational costs.

Significance and Claims

The paper claims that graph-informed NCDEs are particularly beneficial when:

A reliable graph describing flow or interaction structure is available.
The forecasting horizon or spatial resolution requires information to propagate over multiple hops.
Continuous-time evolution of latent trajectories is advantageous.

The authors posit that $INDEQS_{outer}$ serves as a strong, flexible default when performance and adaptability are prioritized, as it combines prior topology with data-driven adjustments. Conversely, $INDEQS_{inner}$ is presented as a compelling alternative when strict adherence to a known physical topology is required and parameter efficiency is a priority. The work highlights the potential of combining neural differential equations with explicit spatial structure, suggesting that leveraging known causal graphs can significantly improve forecasting accuracy over purely data-driven spatial learning, especially in larger, complex systems.