Skip to content

Latest commit

 

History

History
73 lines (66 loc) · 2.79 KB

File metadata and controls

73 lines (66 loc) · 2.79 KB

SHAP-Go Tasks

Completed

v0.3.0

  • TreeSHAP implementation (XGBoost, LightGBM JSON)
  • PermutationSHAP with antithetic sampling
  • SamplingSHAP (Monte Carlo)
  • LinearSHAP - exact closed-form SHAP for linear models
  • KernelSHAP - model-agnostic weighted linear regression (validated against Python SHAP)
  • ExactSHAP - brute-force exact Shapley values (O(n*2^n), validated against mathematical derivations)
  • ONNX Runtime integration
  • Render package (ChartIR visualizations)
  • MkDocs documentation site
  • CI/CD workflows (tests, lint, docs deployment)
  • Benchmarks
  • Fix masker panics (return errors instead)
  • Add ONNX tests
  • Python validation documentation - testing methodology docs
  • Explainers overview page - comparison and decision guide
  • Add internal/rand tests
  • Update README to mark KernelSHAP and ExactSHAP as complete

v0.4.0

  • DeepSHAP - neural network explanations using DeepLIFT rescale rule
  • GradientSHAP - expected gradients using numerical differentiation
  • PartitionSHAP - hierarchical Owen values for feature groupings
  • AdditiveSHAP - exact SHAP for Generalized Additive Models (GAMs)
  • ONNX graph parsing and ActivationSession for intermediate layer capture
  • TreeSHAP interaction values - feature interaction computation
  • CatBoost model support
  • LightGBM text format parser - parse .txt model files
  • ONNX-ML TreeEnsemble parser - native ONNX tree model support
  • Batch explanation API - generic parallel wrapper for any explainer
  • Batched model predictions - WithBatchedPredictions() option for KernelSHAP, ExactSHAP, PartitionSHAP
  • Confidence intervals - uncertainty bounds for sampling methods
  • Improved model/onnx test coverage (15% -> 33%)
  • Background data selection guide - best practices for reference data
  • API reference improvements - comprehensive docs for all explainers
  • Examples (deepshap, gradientshap, multiclass, onnx_basic)
  • Expected output documentation for all examples
  • Udemy-style LMS presentation
  • Coverage badge in README (80% coverage)

Pending

Infrastructure

  • Performance profiling and optimization
  • Memory usage optimization for large models

Coverage Status

Package Coverage
background 90.2%
explainer 58.6%
explainer/additive 89.7%
explainer/deepshap 60.7%
explainer/exact 90.9%
explainer/gradient 87.9%
explainer/kernel 88.5%
explainer/linear 96.6%
explainer/partition 91.7%
explainer/permutation 91.4%
explainer/sampling 93.3%
explainer/tree 84.4%
explanation 77.8%
internal/rand 100.0%
masker 98.0%
model 100.0%
model/onnx 33.2%
render 90.9%
Total 80.6%