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

2026-06-07 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences · Depth: Expert, quick

Summary

A new study investigates the reliability of interaction discoveries in neural time-series models, questioning whether reported variable modulations are data properties or model artifacts. The authors argue this is an identifiability problem, governed by the geometry of observed input support rather than specific neural architecture. Using a multiplicative-gating extension of neural additive vector autoregression (GNAVAR), they demonstrate that representational capacity does not equate to identifiability, as dependent inputs cause leakage and low-dimensional support allows for distinct interaction decompositions. A population identifiability theorem is proven for normalized minimal GNAVAR decompositions under explicit support conditions. This theory yields a practical pre-fit diagnostic: the effective rank of the joint lag-block covariance predicts interaction recovery feasibility. For unknown candidate sets, a two-seed stability check is proposed. The findings, including the identifiability phenomenon and instability signature, are presented as model-agnostic.

Key takeaway

For AI Scientists evaluating neural time-series models for interaction discovery, you must prioritize assessing input data geometry over model complexity. Before fitting, use the effective rank of your joint lag-block covariance as a diagnostic to predict interaction recoverability. If candidate sets are unclear, perform a two-seed stability check across independent model fits. This approach helps you distinguish genuine data interactions from model artifacts, ensuring the reliability of your discoveries.

Key insights

Neural interaction discoveries are identifiable based on input data geometry, not model architecture, with pre-fit diagnostics available.

Principles

• Identifiability is distinct from representational capacity.
• Input support geometry dictates interaction recoverability.
• Instability across fits signals non-identifiable interactions.

Method

Use effective rank of joint lag-block covariance as a pre-fit diagnostic for interaction recovery feasibility. Employ a two-seed stability check for unknown candidate sets.

In practice

• Check effective rank before fitting time-series models.
• Run two independent fits to test interaction stability.
• Evaluate input data geometry for identifiability.

Topics

• Neural Time-Series Models
• Interaction Discovery
• Model Identifiability
• Support Geometry
• GNAVAR
• Pre-Fit Diagnostics

Best for: AI Scientist, Research Scientist

Related on AIssential

• Identifiability of Potentially Degenerate Gaussian Mixture Models With Piecewise Affine Mixing — Identifies latent variables in degenerate GMMs via piecewise affine mixing and sparsity regularization.
• Position: Don't Just "Fix it in Post": A Science of AI Must Study Training Dynamics — A scientific understanding of AI requires studying training dynamics, not just post-training analysis, to predict and control model behaviors.
• I-SAFE: Wasserstein Coherence Metrics for Structural Auditing of Scientific AI Models — Scientific AI model performance benchmarks are insufficient to confirm alignment with domain-relevant structural behavior.
• Identifying Information from Observations with Uncertainty and Novelty — Identification and sample complexity are unified across computational, statistical, and PAC-Bayes learning frameworks.
• Interpreting FCDNNs via RG on Exponential Family — The training of FCDNNs is equivalent to Renormalization Group calculations, explaining their feature extraction and performance.
• Generalized Functional ANOVA in Closed-Form: A Unified View of Additive Explanations — A new method provides a tractable, explicit functional ANOVA decomposition for dependent continuous inputs, enhancing model interpretability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.