wehub-resource-sync
154 lines
4.0 KiB
Plaintext
154 lines
4.0 KiB
Plaintext
{
|
|
"cells": [
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"# Simple Kernel SHAP\n",
|
|
"\n",
|
|
"This notebook provides a simple brute force version of Kernel SHAP that enumerates the entire $2^M$ sample space. We also compare to the full KernelExplainer implementation. Note that KernelExplainer does a sampling approximation for large values of $M$, but for small values it is exact."
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"### Brute Force Kernel SHAP"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 1,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
" reference = [0. 0. 0. 0.]\n",
|
|
" x = [ 1.62434536 -0.61175641 -0.52817175 -1.07296862]\n",
|
|
"shap_values = [ 0.89146267 -0.43752168 -0.31836259 -0.58464256]\n",
|
|
" base_value = 10.000000000000002\n",
|
|
" sum(phi) = 9.55093584213122\n",
|
|
" f(x) = 9.55093584213122\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"import itertools\n",
|
|
"\n",
|
|
"import numpy as np\n",
|
|
"import scipy.special\n",
|
|
"\n",
|
|
"\n",
|
|
"def powerset(iterable):\n",
|
|
" s = list(iterable)\n",
|
|
" return itertools.chain.from_iterable(itertools.combinations(s, r) for r in range(len(s) + 1))\n",
|
|
"\n",
|
|
"\n",
|
|
"def shapley_kernel(M, s):\n",
|
|
" if s == 0 or s == M:\n",
|
|
" return 10000\n",
|
|
" return (M - 1) / (scipy.special.binom(M, s) * s * (M - s))\n",
|
|
"\n",
|
|
"\n",
|
|
"def f(X):\n",
|
|
" np.random.seed(0)\n",
|
|
" beta = np.random.rand(X.shape[-1])\n",
|
|
" return np.dot(X, beta) + 10\n",
|
|
"\n",
|
|
"\n",
|
|
"def kernel_shap(f, x, reference, M):\n",
|
|
" X = np.zeros((2**M, M + 1))\n",
|
|
" X[:, -1] = 1\n",
|
|
" weights = np.zeros(2**M)\n",
|
|
" V = np.zeros((2**M, M))\n",
|
|
" for i in range(2**M):\n",
|
|
" V[i, :] = reference\n",
|
|
"\n",
|
|
" ws = {}\n",
|
|
" for i, s in enumerate(powerset(range(M))):\n",
|
|
" s = list(s)\n",
|
|
" V[i, s] = x[s]\n",
|
|
" X[i, s] = 1\n",
|
|
" ws[len(s)] = ws.get(len(s), 0) + shapley_kernel(M, len(s))\n",
|
|
" weights[i] = shapley_kernel(M, len(s))\n",
|
|
" y = f(V)\n",
|
|
" wsq = np.sqrt(weights)\n",
|
|
" result = np.linalg.lstsq(wsq[:, None] * X, wsq * y, rcond=None)[0]\n",
|
|
" return result\n",
|
|
"\n",
|
|
"\n",
|
|
"M = 4\n",
|
|
"np.random.seed(1)\n",
|
|
"x = np.random.randn(M)\n",
|
|
"reference = np.zeros(M)\n",
|
|
"phi = kernel_shap(f, x, reference, M)\n",
|
|
"base_value = phi[-1]\n",
|
|
"shap_values = phi[:-1]\n",
|
|
"\n",
|
|
"print(\" reference =\", reference)\n",
|
|
"print(\" x =\", x)\n",
|
|
"print(\"shap_values =\", shap_values)\n",
|
|
"print(\" base_value =\", base_value)\n",
|
|
"print(\" sum(phi) =\", np.sum(phi))\n",
|
|
"print(\" f(x) =\", f(x))"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"### Using KernelExplainer"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 2,
|
|
"metadata": {
|
|
"scrolled": false
|
|
},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"shap_values = [ 0.89146267 -0.43752168 -0.31836259 -0.58464256]\n",
|
|
"base value = 10.0\n"
|
|
]
|
|
}
|
|
],
|
|
"source": [
|
|
"import shap\n",
|
|
"\n",
|
|
"explainer = shap.KernelExplainer(f, np.reshape(reference, (1, len(reference))))\n",
|
|
"shap_values = explainer.shap_values(x)\n",
|
|
"print(\"shap_values =\", shap_values)\n",
|
|
"print(\"base value =\", explainer.expected_value)"
|
|
]
|
|
}
|
|
],
|
|
"metadata": {
|
|
"anaconda-cloud": {},
|
|
"kernelspec": {
|
|
"display_name": "Python 3",
|
|
"language": "python",
|
|
"name": "python3"
|
|
},
|
|
"language_info": {
|
|
"codemirror_mode": {
|
|
"name": "ipython",
|
|
"version": 3
|
|
},
|
|
"file_extension": ".py",
|
|
"mimetype": "text/x-python",
|
|
"name": "python",
|
|
"nbconvert_exporter": "python",
|
|
"pygments_lexer": "ipython3",
|
|
"version": "3.9.16"
|
|
}
|
|
},
|
|
"nbformat": 4,
|
|
"nbformat_minor": 1
|
|
}
|