📊 statistics

When Are Neural Interaction Discoveries Real? Identifiability, Recoverability, and a Pre-Fit Diagnostic

This paper establishes that the identifiability of neural interaction discoveries depends fundamentally on the geometry of the observed input support rather than model architecture, and it introduces a pre-fit diagnostic based on effective rank and a stability check to determine whether such interactions can be reliably recovered.

Original authors: Valentina Kuskova, Dmitry Zaytsev, Michael Coppedge

Published 2026-06-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Valentina Kuskova, Dmitry Zaytsev, Michael Coppedge

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

The Big Question: Is the AI "Seeing" or Just "Guessing"?

Imagine you hire a very smart, flexible detective (a neural network) to figure out how a crime happened. The detective tells you: "The suspect's speed was modulated by the weather." In other words, the weather made the suspect run faster.

This sounds like a great, mechanistic discovery. But the authors of this paper ask a crucial question: Is this actually a fact about the world, or is it just a trick of the detective's flexible mind?

Because these AI models are so good at fitting data, they can often invent different stories that all look exactly the same on the surface. One story might say "Weather caused the speed," while another says "The suspect's mood caused the speed," and both stories fit the historical data perfectly. If the model picks one story just because of how it started its calculation (a random "seed"), that story isn't a real discovery; it's an artifact.

The paper argues that whether a discovery is real depends less on how smart the AI is, and more on the shape of the data it was fed.

The Core Concept: The "Stage" vs. The "Actors"

The authors use a specific type of AI model called G-NAVAR (Gated Neural Additive Vector Autoregression). Think of this model as a stage play:

The Actors (Sources): Variables that influence the outcome (e.g., wind speed, temperature).
The Director (Gates): Variables that change how the actors perform (e.g., "When it's hot, the wind blows harder").

The paper asks: Can we uniquely identify who the Director is?

1. The Problem of "Leaking" (Dependent Inputs)

Imagine two actors, Rain and Clouds, who always show up together. If the model tries to figure out who is directing the scene, it gets confused. Because Rain and Clouds are so linked, the model can't tell if the "Director" is Rain or Clouds. The effect "leaks" between them.

The Paper's Claim: If your input variables are too dependent on each other, the model cannot uniquely separate their effects.

2. The Problem of the "Small Stage" (Low-Dimensional Support)

This is the paper's most important finding. Imagine a stage that is only 1 inch wide. You have two actors trying to perform a complex dance. Because the stage is so narrow, they can only move in one direction.

The Analogy: If the data only covers a tiny, flat slice of reality (like a straight line), the model can invent a million different "dance moves" (interaction rules) that all look identical on that tiny line. But if you stepped off the line into the rest of the room, the moves would look totally different.
The Paper's Claim: If the data doesn't cover enough "ground" (geometric support), the model's discovery is undefined. It's not a real property of the data; it's just a guess that happens to work on the narrow slice of data you have.

The Solution: A "Pre-Flight" Check and a "Double-Check"

The authors propose a practical workflow to stop people from trusting fake discoveries. They offer two tools:

Tool 1: The "Effective Rank" (The Pre-Fit Diagnostic)

Before you even train the AI, you can look at your data and ask: "Is our stage wide enough?"

They use a math trick called Effective Rank. Think of it like checking how many dimensions your data actually uses.
High Effective Rank: Your data is spread out in many directions (a big, 3D room). The model might be able to find a real interaction.
Low Effective Rank: Your data is squashed into a flat sheet or a line (a narrow hallway). The model cannot find a real interaction, no matter how hard it tries.
The Rule: If the Effective Rank is low, stop immediately. The discovery is impossible.

Tool 2: The "Two-Seed" Stability Check (The Post-Fit Test)

If the "stage" looks wide enough, you still need to be careful. You train the model twice, starting with two different random "seeds" (like rolling two different dice to start the game).

The Test: Do both runs agree on who the Director is?
If they agree: It's a good sign (though not a guarantee).
If they disagree: The discovery is fake. The model is just picking a random answer because the data doesn't force a single truth.

Real-World Tests: Three Different Stories

The authors tested their theory on three real-world datasets, which perfectly illustrate the three possible outcomes:

1. Beijing Air Quality (The "Success" Story)

The Data: Air pollution and weather.
The Result: The data was "rich" (High Effective Rank). The model consistently found that Temperature modulates the relationship between Nitrogen Dioxide and Ozone.
Verdict: This is a Real Discovery. The data was wide enough, and the model agreed on the answer every time.

2. World Development Indicators (The "False Hope" Story)

The Data: Economic growth and investment.
The Result: The data looked "rich" (High Effective Rank), so the Pre-Fit check said "Go ahead!" But when they ran the model twice, the two runs disagreed. One said "Investment" was the key; the other said "Trade Openness."
Verdict: This is Unstable. Even though the data looked good, there was no single, recoverable truth. The model was just guessing.

3. Stock Market Volatility (The "Impossible" Story)

The Data: Stock market prices across different countries.
The Result: The data was "collapsed" (Low Effective Rank). All the stock markets moved in lockstep, like a single line. The Pre-Fit check screamed "Stop!"
Verdict: Impossible. The model couldn't find a real interaction because the data was too narrow. Any "discovery" would be pure fiction.

The Bottom Line

The paper concludes that having a flexible AI model is not enough. You can have the most powerful neural network in the world, but if your data is "squashed" (low-dimensional) or too correlated, the model cannot tell you the truth about how variables interact.

The Takeaway for Practitioners:

Check the shape of your data first (Effective Rank). If it's too narrow, don't bother looking for interactions.
Run the model twice. If the answers change, you haven't found a truth; you've found a random guess.
Don't trust the "story" the AI tells you unless the data geometry and the stability checks both say it's real.

The paper doesn't promise that AI will solve all mysteries; it promises to tell you when the AI is actually solving a mystery and when it is just making things up.

Technical Summary: When Are Neural Interaction Discoveries Real?

Problem Statement

The paper addresses a fundamental question in interpretable time-series modeling: When a neural model (specifically one designed to capture interactions) reports that one variable modulates the effect of another on a target, is this discovered interaction a genuine property of the data-generating process, or merely an artifact of the model's flexibility?

While neural additive vector autoregression (NAVAR) and related Granger causality methods can decompose dynamics into interpretable source contributions, they cannot naturally express effect modification (where the influence of one variable depends on the level of another). Extending these models to include multiplicative gates (G-NAVAR) allows for such interactions, but the authors argue that representational capacity does not guarantee identifiability. A flexible model may admit multiple, observationally equivalent interaction decompositions that fit the data equally well but differ in their structural interpretation. The central problem is determining whether the recovered interaction structure is uniquely identifiable from observational data, or if it is an optimization artifact driven by the geometry of the input support.

Methodology

Model Architecture: G-NAVAR

The authors study Gated Neural Additive Vector Autoregression (G-NAVAR). For a target variable $x_i$ , the model decomposes the conditional mean into a sum of edge products:
$\hat{x}_{i,t} = \beta_i + \sum_j f_{ij}(x^{(K)}_{j,t}) \prod_{k \in M_{ij}} g_{ijk}(x^{(K)}_{k,t})$
Here, $f_{ij}$ is a base function for the source $j$ , and $g_{ijk}$ are strictly positive multiplicative gates driven by candidate modulators $k \in M_{ij}$ . The product $F_{ij} = f_{ij} \prod g_{ijk}$ represents the effective contribution of source $j$ .

Theoretical Framework

The paper establishes identifiability conditions by analyzing gauge symmetries and support geometry:

Normalization and Minimality: To remove trivial symmetries (scale redistribution, permutation of modulators, insertion of gates equal to 1), the authors impose:
- Normalization: $E[g_{ijk}] = 1$ .
- Minimality: Only essential modulators (those that non-trivially change the edge product) are included.
Obstructions to Identifiability:
- Dependence-Induced Leakage: If inputs are dependent, conditioning does not isolate source effects; lower-order variations can leak between edge-specific terms.
- Low-Dimensional Support: If the joint support of modulators is effectively low-dimensional (e.g., collinear), distinct gate functions can agree perfectly on the observed data while differing elsewhere, making recovery impossible.
Identifiability Theorem: The authors prove a population identifiability theorem for normalized minimal G-NAVAR under specific conditions:
- Edge Separability: Contributions from different sources must be distinguishable. This is achieved via a direct-sum condition for disjoint supports or a Hierarchically Orthogonal Functional Decomposition (HOFD) for shared modulators.
- Support Richness: The joint support of the modulator set must be a full-dimensional, open, connected product set with a strictly positive density.
- Gate-Interaction Faithfulness: The listed modulator set must correspond to the structural interaction components.

Practical Diagnostic Tools

The theory yields two operational tools for practitioners:

Pre-Fit Diagnostic (Effective Rank): The effective rank ( $r_{eff}$ ) of the joint lag-block covariance matrix serves as a proxy for support richness.
$r_{eff}(S) = \frac{(\text{tr } \Sigma_S)^2}{\text{tr}(\Sigma_S^2)}$
A low effective rank indicates support collapse (degeneracy), predicting non-identifiability before any model fitting occurs.
Post-Fit Stability Check: A two-seed stability check involves fitting the model twice from independent random seeds. If the recovered modulator rankings disagree significantly, the interaction structure is likely not stably recoverable, even if the support appears rich.

Key Results

Synthetic Experiments

Recovery at Scale: In a synthetic setting with rich support, G-NAVAR successfully recovers the true modulator sets and gate shapes as sample size increases, outperforming additive baselines (which cannot capture multiplicative structure) and matching the predictive accuracy of black-box MLPs and functional ANOVA models (GA2M).
Support Collapse Transition: As the correlation between modulators increases (reducing effective rank), the recovery of modulator sets collapses. The error in gate estimation rises continuously as $r_{eff}$ decreases, confirming that identifiability is a function of support geometry, not just sample size.
Model Agnosticism: Under rich support, multiple interaction-aware models (G-NAVAR, GA2M) recover the structure, suggesting that the parameterization is less critical than the support condition for recovery.

Real-Data Experiments

The authors validate the theory across three distinct empirical regimes:

Beijing Air Quality (Rich Support, Recoverable):
- Context: Modeling $NO_2 \to O_3$ and $PM_{2.5} \to PM_{10}$ edges.
- Result: High effective rank ( $r_{eff} > 4$ ) coincides with stable recovery of the expected modulators (Temperature for ozone, Wind Speed for dispersion) across all sites. The pre-fit diagnostic correctly predicted recoverability.
WDI Development Indicators (Rich Support, Unstable):
- Context: Modeling the effect of resource rents on GDP growth modulated by investment/trade.
- Result: Despite high effective rank ( $r_{eff} \approx 4.47$ ), the two-seed check revealed instability (disagreement on top modulators, sign flips in gate slopes). This demonstrates that rich support is necessary but not sufficient; the interaction structure itself may be weak or unidentifiable in the data distribution.
Realized Volatility (Collapsed Support, Unrecoverable):
- Context: Modeling volatility spillovers between global equity indices.
- Result: Extremely low effective rank ( $r_{eff} < 2$ ) due to high cross-correlation among markets. The pre-fit diagnostic correctly flagged the regime as non-identifiable. Recovery was unstable, and the model failed to outperform additive baselines.

Significance and Claims

The paper claims that interaction discoverability is fundamentally a problem of identifiability governed by the geometry of the observed input support, rather than the specific neural architecture used.

Distinction between Capacity and Identifiability: A model's ability to represent interactions does not imply it can recover them uniquely from data.
Model-Agnostic Principles: While the formal proofs are derived for G-NAVAR, the authors argue that the phenomena of support-dependent non-identifiability and the necessity of stability checks apply to any flexible neural autoregressive method attempting to discover interactions.
Practical Workflow: The paper provides a concrete, two-step workflow for practitioners:
1. Compute the effective rank of the candidate modulator set before fitting. If low, recovery is impossible regardless of model complexity.
2. If the rank is adequate, perform a two-seed stability check. If the results are unstable, the discovered interaction is likely an artifact of optimization rather than a data property.
Limitations: The authors are modest in their claims, noting that the diagnostic is necessary but not sufficient (it cannot detect higher-order misspecification) and that the "stability" of a result does not guarantee correctness, only reproducibility.

In summary, the paper shifts the focus from "can we fit an interaction?" to "is the interaction uniquely determined by the data?" providing both a theoretical foundation and practical tools to distinguish real discoveries from optimization artifacts.