A new research framework shows that knowing why an AI model makes a prediction matters just as much as knowing it's correct. Researchers at Backwell Tech have published a study introducing IMEX, an interaction-based explainability method that reveals not just which variables influence a prediction, but how combinations of variables work together to shape the outcome.

Why Does AI Explainability Matter in Real Business?

Black-box AI models can be remarkably accurate without being transparent. They produce the right answer without showing the mechanism behind it, which creates a trust problem in high-stakes contexts like supply chain planning, pricing decisions, or operational forecasting. A plant manager or pricing lead needs to understand not just what the model predicts, but why, so they can act on that prediction with confidence. This is where explainability becomes essential.

Emiliano Massi, Head of Research and Development at Backwell Tech, published the IMEX paper on arXiv, introducing a post-hoc explainability method that works on top of already-trained models like XGBoost rather than requiring a completely different architecture. Unlike many existing explainability approaches, IMEX can investigate how three, four, or more variables interact together, not just pairs.

How Does IMEX Reveal Hidden Variable Interactions?

The framework rests on two complementary metrics designed to build an interpretability map for a model's predictions:

• Static Correlation Power (PCS): Quantifies each individual feature's contribution to a prediction, measuring how much each variable matters on its own.
• Interaction Correlation Power (PCI): Captures the non-additive effects between features, revealing how variables amplify or dampen each other's influence.
• Higher-Order Interactions: Extends analysis beyond pairs to show how three, four, or more variables work together in ways that simple addition cannot explain.

In the paper's experiments, researchers benchmarked PCS against INVASE, an established feature-selection method, across three synthetic datasets modeled on real-world scenarios: retail purchasing behavior, online purchases, and fuel station operations. The results showed that PCS recovered relevant feature-level structure even where relationships between inputs and outcomes were non-linear, conditional, or multicollinear, and behaved more consistently than INVASE in several conditions.

"A prediction is only as useful as the reasons behind it. And truthfully, explainability shouldn't be a black box either," stated Emiliano Massi, Head of Research and Development at Backwell Tech.

Emiliano Massi, Head of Research and Development at Backwell Tech

What Makes This Approach Different From Existing Methods?

The paper is careful to frame IMEX as complementary rather than a simple replacement for existing methods. Depending on the scenario, IMEX and INVASE sometimes agree, sometimes complement each other, or pick up on different parts of the same structure. This nuance matters because it means practitioners can use multiple explainability tools together rather than betting everything on a single approach.

The current paper focuses on empirical validation of PCS, with PCI and the extension to higher-order interactions introduced as methodological contributions. The authors have set out dedicated empirical validation of these extensions as future work, meaning the full power of the framework is still being tested.

Steps to Implement Explainability in Your AI Models

• Assess Your Use Case: Determine whether your AI predictions affect high-stakes decisions like pricing, supply chain, or operations where understanding the reasoning is critical for trust and compliance.
• Choose Post-Hoc Methods: Consider explainability frameworks that work on top of existing trained models rather than requiring you to rebuild your entire architecture from scratch.
• Test Multiple Approaches: Use complementary explainability methods together to cross-validate findings, since different techniques can reveal different aspects of how your model works.
• Validate Against Known Structures: When possible, test your explainability method on datasets where you already know the ground-truth relationships between variables, to ensure the method is reliable.

The timing of this research reflects a broader shift in how companies think about AI. Accuracy alone is no longer sufficient; stakeholders increasingly demand transparency. Regulators, customers, and internal teams want to know not just that an AI system works, but how it works. This is especially true in industries like food and fuel distribution, where Backwell Tech focuses its work, where supply chain decisions ripple across entire networks.

The IMEX paper is available on arXiv for researchers and practitioners interested in exploring the full technical details and experimental results. For companies already using predictive AI models, the framework offers a practical pathway to add explainability without overhauling existing systems.