Expand description
§Interpretability Tools
Comprehensive model interpretability toolkit including SHAP integration, LIME support, attention analysis, feature attribution, and counterfactual generation for TrustformeRS models.
§Refactoring Summary
Previously this was a single 2,803-line file containing all interpretability functionality. It has been split into focused modules:
interpretability/config.rs- Configuration structures and enums (77 lines)interpretability/shap.rs- SHAP analysis types and functionality (66 lines)interpretability/lime.rs- LIME analysis types and functionality (78 lines)interpretability/attention.rs- Attention analysis for transformers (426 lines)interpretability/attribution.rs- Feature attribution methods (103 lines)interpretability/counterfactual.rs- Counterfactual generation (191 lines)interpretability/analyzer.rs- Main analyzer implementation (318 lines)interpretability/report.rs- Reporting functionality (23 lines)
This refactoring improves:
- Code maintainability and readability
- Module compilation times
- Test isolation
- Code reuse through focused modules
- Developer experience when working on specific interpretability methods
Structs§
- Actionable
Insight - Actionable insight from counterfactual analysis
- Attention
Analysis Result - Attention analysis result for transformer models
- Attention
Bottleneck - Attention bottleneck
- Attention
Flow Analysis - Attention flow analysis
- Attention
Flow Path - Attention flow path
- Attention
Head Result - Attention analysis for a single head
- Attention
Insight - Attention insight
- Attention
Layer Result - Attention analysis for a single layer
- Attention
Patterns - Overall attention patterns
- Attention
Statistics - Attention statistics
- Attribution
Method Result - Attribution result for a specific method
- Attribution
Visualization Data - Data for visualizing attributions
- Block
Pattern - Block attention pattern (sequence segments)
- Boundary
Crossing Point - Point where instance crosses decision boundary
- Counterfactual
- Individual counterfactual example
- Counterfactual
Quality Metrics - Quality metrics for counterfactuals
- Counterfactual
Result - Counterfactual generation result
- Decision
Boundary Analysis - Decision boundary analysis
- Diagonal
Pattern - Diagonal attention pattern (local dependencies)
- Feature
Attribution - Individual feature attribution
- Feature
Attribution Result - Feature attribution analysis result
- Feature
Change - Change made to a feature in counterfactual
- Feature
Contribution - Individual feature contribution
- Feature
Importance - Feature importance from LIME
- Feature
Interaction - Feature interaction information
- Feature
Sensitivity Analysis - Feature sensitivity analysis from counterfactuals
- Flow
Efficiency Metrics - Flow efficiency metrics
- Head
Cluster - Cluster of similar attention heads
- Head
Redundancy Analysis - Head redundancy analysis
- Head
Specialization Analysis - Head specialization analysis
- Interaction
Effect - Effect of feature interactions on sensitivity
- Interpretability
Analyzer - Main interpretability analyzer
- Interpretability
Config - Configuration for interpretability tools
- Interpretability
Report - Comprehensive interpretability report
- Layer
Attention Patterns - Layer-level attention patterns
- Layer
Attention Stats - Layer attention statistics
- Layer
Flow Stats - Layer flow statistics
- Layer
Flow Step - Flow step through a layer
- Lime
Analysis Result - LIME (Local Interpretable Model-agnostic Explanations) analysis result
- Method
Agreement Analysis - Analysis of agreement between attribution methods
- Neighborhood
Stats - Local neighborhood statistics
- Perturbation
Analysis - Perturbation analysis details
- Perturbation
Result - Individual perturbation result
- Pruning
Impact - Expected impact of pruning a head
- Pruning
Recommendation - Recommendation for head pruning
- Redundant
Head Pair - Pair of redundant attention heads
- Repetitive
Pattern - Repetitive attention pattern
- Shap
Analysis Result - SHAP (SHapley Additive exPlanations) analysis result
- Shap
Summary - SHAP analysis summary
- Sparsity
Distribution - Sparsity distribution across layers and heads
- Specialization
Evolution - Evolution of specialization across layers
- Specialization
Transition - Transition between specialization types
- Threshold
Analysis - Analysis of decision thresholds
- Timeline
Point - Point in attribution timeline
- Token
Attention Score - Token attention information
- TopFeature
- Top attributed feature with additional information
- Vertical
Pattern - Vertical attention pattern (specific token focus)
Enums§
- Attribution
Method - Attribution methods for feature importance
- Bottleneck
Type - Type of attention bottleneck
- Change
Direction - Direction of feature change
- Flow
Transformation - Type of flow transformation
- Head
Specialization Type - Types of attention head specialization
- Implementation
Difficulty - Difficulty of implementing suggested changes
- Insight
Type - Type of attention insight
- Interaction
Effect Type - Type of interaction effect
- Interaction
Type - Type of feature interaction
- Redundancy
Type - Type of redundancy between heads
- Risk
Level - Risk level for pruning
- Specialization
Trend - Overall specialization trend
- Time
Horizon - Time horizon for implementing changes