pulsatrix::KernelSHAP Class Reference

Approximates Shapley values via full coalition enumeration + SHAP-kernel-weighted linear regression, reusing fit_weighted_linear_regression for the reduced (n-1)-dimensional problem the efficiency-axiom substitution produces. More...

#include <kernel_shap.hpp>

Public Member Functions
Attribution	explain (const std::function< Tensor(const Tensor &)> &predict, const Tensor &input, const Tensor &baseline, int64_t target_index, DeviceBackend *backend) const
	Computes Shapley value approximations for one output index.

Detailed Description

Approximates Shapley values via full coalition enumeration + SHAP-kernel-weighted linear regression, reusing fit_weighted_linear_regression for the reduced (n-1)-dimensional problem the efficiency-axiom substitution produces.

Note

The empty/full coalitions have undefined SHAP kernel weight (division by zero) – phi_0 is fixed to f(baseline) exogenously, and the efficiency axiom (sum(phi_i) = f(x) - f(baseline)) is enforced by substitution (eliminating the last feature's coefficient), not by including those two coalitions in the regression. See mission_kernel_shap.md's Recon for the hand-derived n=2 case this was verified against before implementation.

Full 2^n enumeration, not sampling – exact at this codebase's small feature counts (mathematically equivalent to the definitional Shapley formula at full enumeration), not the charter's excluded O(2^n)/O(n!) brute-force-at-scale case.

Takes a forward-pass callable, not ExplainerContext& – model-agnosticism by construction, same pattern as LIME.

Device-generic host boundary: coalitions are built on the host from one read of input/baseline, uploaded beside the input, and only f(z)[target] is read back.

Member Function Documentation

explain()

Attribution pulsatrix::KernelSHAP::explain ( const std::function< Tensor(const Tensor &)> &  predict, const Tensor &  input, const Tensor &  baseline, int64_t  target_index, DeviceBackend *  backend  ) const

inline

Computes Shapley value approximations for one output index.

Parameters

predict	Forward-pass callable: input Tensor in, output Tensor out.
input	Input to explain.
baseline	Reference "feature absent" input. Must match input's shape.
target_index	Which output element to attribute (0-based, flat index).
backend	Backend to allocate intermediate tensors through.

Returns

An Attribution with method "kernel_shap", values = the per-feature Shapley values (same shape as input), and metadata recording the target index used.

---

The documentation for this class was generated from the following file:

• include/pulsatrix/kernel_shap.hpp