Quantifying Synergistic Feature Interactions in Black Box Models

This direction is split into other directions: feature independence assumption (feature interaction neglect). Its last version stays readable and citable here.

Version 1 of 1106 papers119 conceptsfrozen 2026-10-04

Most interpretability methods treat machine learning models as a collection of independent voices. If a model predicts a high risk of heart disease, tools like SHAP or LIME typically assign a "weight" to age and a "weight" to blood pressure as if they were separate factors. However, in reality, age and blood pressure are often only predictive when considered together. This research direction focuses on moving beyond "flat" feature importance to capture synergy—the predictive power that emerges only through the interaction of multiple variables.

The Failure of Univariate Attribution

The core problem identified across this research is the "independence assumption." Traditional feature attribution methods evaluate the contribution of a feature by perturbing it in isolation. As Cheng-Yu Hsieh and Pradeep Ravikumar argue in Evaluations and Methods for Explanation through Robustness Analysis, this fails to capture cases where features are only significant as a set.

Consider a simple XOR logic gate: if two inputs are the same, the output is 0; if they differ, the output is 1. If you look at either input alone, it tells you nothing about the output. A univariate attribution method would assign both features a score of zero, even though they are 100% responsible for the result. Nils Strodthoff and colleagues demonstrate in PredDiff: Explanations and Interactions from Conditional Expectations that this phenomenon hides the true drivers of decision-making in complex classifiers.

Measuring Joint Importance with Game Theory

To fix this, researchers are extending cooperative game theory—specifically Shapley values—to account for groups. While the standard Shapley value tells you the average marginal contribution of a single player (feature), new frameworks like the Shapley-Taylor index, introduced by Mukund Sundararajan and others in A New Interaction Index inspired by the Taylor Series, provide a way to attribute credit to interactions.

The intuition here is to treat a pair or triplet of features as a single "player" in the game. Chris Harris and Richard Pymar propose Joint Shapley values to measure the total contribution of a feature subset as a cohesive unit. This prevents the "dilution" of importance that happens when a synergistic effect is forced back into individual components.

The Credit Assignment Problem

If a pair of features produces a "synergy bonus," how should that bonus be distributed? This is the central mathematical challenge of the field. Daniel Lundstrom and Meisam Razaviyayn address this in A Unifying Framework to the Analysis of Interaction Methods using Synergy Functions. They show that different XAI methods are essentially different ways of partitioning these higher-order effects.

For example, in a medical diagnosis, "Age" and "Smoking History" might interact. A method might:

  1. Assign the interaction's value entirely to a "joint" score.
  2. Split it 50/50 between the two features.
  3. Weight the distribution based on the functional influence (e.g., polynomial degree) of each feature.

Gunnar König and Eric Günther propose the DIP (Disentangling Interactions and Dependencies) framework in Disentangling Interactions and Dependencies in Feature Attribution to formally separate standalone contributions from these shared synergies.

Beyond Attribution: Selection and Logic

This shift isn't just about explaining existing models; it's about building better ones. In medical data mining, Sam Chao and Qiulin Ding developed techniques like LUIFS to rescue "useless" features that have latent supportive relevance—variables that fail univariate tests but are critical when paired. Similarly, Chohee Kim and Mihaela van der Schaar introduced SynFS to find synergistic interactions across different data "views," such as genetics and clinical records.

In Natural Language Processing, the interaction is often the message itself. Chen Qian developed CIDR, which treats word interactions as a knapsack optimization problem to find the minimal set of words that drive a prediction. This captures linguistic effects like negation (e.g., "not" + "good"), where the meaning of one word is entirely contingent on the other.

Scaling to High Dimensions

The computational cost of calculating all possible interactions is exponential. To make these insights practical, researchers like Fabian Fumagalli and Maximilian Muschalik have developed SHAP-IQ and the shapiq library. These tools use efficient sampling and stratification to approximate any-order interactions, allowing practitioners to explain complex models like Vision Transformers or LLMs without waiting for an exhaustive combinatorial search.

By moving from "what features are important" to "how do features work together," this research direction provides a more faithful map of the logic inside modern black-box systems.

Papers in this version

Showing 60 of 106 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_quantifying_synergistic_feature_interact_77f7184b,
  title = {Quantifying Synergistic Feature Interactions in Black Box Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/quantifying-synergistic-feature-interactions-in-black-box-models/rdr_77f7184b638d41dbb45fd648c28396e8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_quantifying_synergistic_feature_interact_77f7184b_v1,
  title = {Quantifying Synergistic Feature Interactions in Black Box Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/quantifying-synergistic-feature-interactions-in-black-box-models/rdr_77f7184b638d41dbb45fd648c28396e8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}