Feature Attribution in Directed Acyclic Graphs Using Edge Intervention

2026-06-13 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Mathematics & Computational Sciences · Depth: Expert, quick

Summary

DAG-SHAP is a novel feature attribution method designed for Directed Acyclic Graphs (DAGs) that addresses limitations of existing Shapley value-based approaches. Traditional methods often adopt a node-centric view, failing to capture the externality and exogenous influence of features in complex causal relationships. Proposed by the authors, DAG-SHAP employs an edge intervention strategy, treating each feature edge as an individual attribution object. This ensures that both the externality and exogenous contributions of features are appropriately captured. The method also includes an approximation technique for efficient computation. Its effectiveness has been validated through extensive experiments on both real and synthetic datasets. The code for DAG-SHAP was made available on GitHub, and the paper was published on 2026-06-13.

Key takeaway

For Machine Learning Engineers or Data Scientists working with models on Directed Acyclic Graphs, if you need more accurate feature attribution than traditional Shapley values provide, consider implementing DAG-SHAP. Its edge intervention approach better captures complex causal relationships and exogenous influences, offering more reasonable interpretations. You should explore the provided GitHub repository to integrate this method for improved model explainability.

Key insights

DAG-SHAP attributes feature importance in DAGs by intervening on edges, capturing complex causal influences beyond node-centric views.

Principles

• Node-centric attribution can misinterpret complex causal features.
• Edge intervention captures externality and exogenous influence.
• Feature attribution benefits from a causal structure view.

Method

DAG-SHAP attributes importance by treating each feature edge in a DAG as an individual attribution object, using an edge intervention approach. An approximation method enables efficient computation.

In practice

• Apply DAG-SHAP for causal feature importance in DAGs.
• Use the approximation for large-scale DAG models.
• Explore code on GitHub for implementation.

Topics

• Feature Attribution
• Shapley Values
• Directed Acyclic Graphs
• Causal Inference
• Edge Intervention
• Model Explainability

Code references

• ZJU-DIVER/DAG-SHAP

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Data Scientist

Related on AIssential

• TN-SHAP-G: Graph-Structured Tensor Network Surrogates for Shapley Values and Interactions — TN-SHAP-G efficiently computes Shapley values for graph-structured inputs using a graph-aligned tensor network surrogate.
• Towards Rigorous Explainability by Feature Attribution — Existing SHAP scores for ML explainability are flawed and misleading; rigorous alternatives are necessary.
• ADAG: Automatically Describing Attribution Graphs — Automated circuit tracing and interpretation can reveal LLM internal computations and identify steerable features.
• Interpretable Factor Decomposition for Decision Intelligence in Large-Scale Financial Markets: Evidence from China's A-Share Market — Interpretable ML reveals behavioral factors dominate Chinese A-share predictability, offering auditable financial insights.
• Forecasting Is Not Attribution: Localizing Decoder Bypass in Graph-Based Neural Marketing Mix Models — Forecasting accuracy in neural MMMs does not guarantee correct attribution, often hiding a "decoder bypass" failure.
• The Attribution Impossibility: No Feature Ranking Is Faithful, Stable, and Complete Under Collinearity — Feature ranking cannot be simultaneously faithful, stable, and complete under collinearity.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.