At Backwell, we've built our product on a simple premise: a prediction is only as useful as the reasons behind it. And truthfully, explainability shouldn't be a black box either; this premise now has an academic contribution behind it too.
Emiliano Massi, Backwell Tech's Head of Research and Development, has published "IMEX: Interaction-Based Model Explanation" on arXiv, introducing a new approach to explainable predictive modeling.
Black-box models can be accurate without being transparent. They can produce the right answer without showing the mechanism behind it, and in high-stakes contexts, accuracy alone isn't a sufficient basis for trust. IMEX addresses this as a post-hoc explainability method: it works on top of an already-trained model (XGBoost, in the paper's experiments) rather than requiring a different model architecture. It identifies which variables matter most to a prediction, and, critically, which combinations of variables interact to shape it. Unlike many explainability methods, IMEX places no limit on the order of these interactions: it can investigate how three, four, or more variables work together, not just pairs.
The framework rests on two complementary metrics: Static Correlation Power (PCS), which quantifies each feature's individual contribution, and Interaction Correlation Power (PCI), which captures the non-additive effects between features. Together, they're designed to build toward an interpretability map for a model's predictions. In this paper, PCS is the metric put through empirical testing: it's benchmarked against INVASE, an established instance-wise feature-selection method, across three synthetic datasets with known ground-truth structures, modeled on retail purchasing behavior, online purchases, and fuel station operations. The results show PCS recovering relevant feature-level structure even where relationships between inputs and outcomes are non-linear, conditional, or multicollinear, and behaving more consistently than INVASE in a number of those conditions. The paper is careful to frame this as complementary rather than a simple win: depending on the scenario, the two methods agree, complement each other, or pick up on different parts of the same structure. PCI and the extension to higher-order interactions are introduced as principled methodological contributions in this paper, with their own dedicated empirical validation set out as future work.
It's the same thinking that runs through how we build predictive modeling capability at Backwell more broadly: knowing that a model is right is not enough. Knowing why is what lets the person acting on it, a plant manager, a pricing lead, a distribution planner, actually trust and use the result. It's a fitting detail that one of the paper's own test cases models a fuel station's operations, close to home for a company built around food and fuel distributors.
Read the full paper on arXiv: https://arxiv.org/abs/2607.14096
About Backwell Tech
Backwell Tech is a Berlin-based high-tech company specializing in predictive AI solutions. The platform offers companies scalable AI models for profit maximization by utilizing historical and real-time data and ensuring data integrity. Since its founding in 2019, Backwell Tech has combined cutting-edge research with practical innovation in explainable algorithms. The company focuses on ethical AI development and delivers reliable, interpretable forecasts that enable informed business decisions. More information at www.backwelltechcorp.com.
Backwell Tech Corp contact:
Maximilian Gismondi
hello@backwelltechcorp.com