Home / Curated / Algorithms / Deep SHAP

Deep SHAP

Algorithms AIMech:DeepSHAP DRAFT

A compositional approximation of neural-network Shapley attributions using component-level SHAP values and backward multipliers.

Identity Notes & Concept Scope Deep SHAP adapts DeepLIFT-style propagation with component SHAP values; it is a distinct approximation method, not the general SHAP definition or a synonym for all DeepLIFT rules.

Claims & Grounded Evidence

1 Claim

Mechanism #explanation-method

Compute attributions for supported simple network components, convert them to multipliers relative to reference expectations, and compose those multipliers backward to approximate whole-network feature attributions.

supports Source: A Unified Approach to Interpreting Model Predictions · Locator: Section 4.2, Deep SHAP, equations 13–16 and Figure 2B, PDF/printed p. 7

⚠️ Curated Limitations & Negative Results

Component-wise exact values do not imply exact Shapley values for an arbitrary composite nonlinear model.
Reference/background and feature-dependence assumptions affect the approximation; no causal-importance guarantee is curated.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:ArtificialNeuralNetwork` Artificial Neural Network	related_match	The method directly propagates attributions through artificial neural networks. The mapping relates an analysis procedure to its analyzed system, not to an equivalent architecture.

Evidence Sources & Primary Literature

1 Source

Source ID	Publication / Document	Type / Access	Reference & Checksum
`original-paper`	A Unified Approach to Interpreting Model Predictions Scott M. Lundberg, Su-In Lee	primary paper full_text	sha256:32518459c313…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-11T04:02:04Z

Updated: 2026-10-11T04:02:04Z

Source File: data/curated/DeepSHAP.yaml