File size: 16,258 Bytes
b9f807b 8078e25 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 | """Metrics computation for unlearning experiments."""
import numpy as np
from scipy import stats
from scipy.special import digamma, polygamma
from typing import Dict, List, Tuple, Optional
from collections import defaultdict
from src.graph_utils import build_adjacency, get_deletion_neighborhood, get_blocks_at_distance
# ============================================================
# Parameter distance / influence
# ============================================================
def compute_block_param_vector(params: dict, node_type: str, idx: int, model_family: str) -> np.ndarray:
"""Extract flat parameter vector for a single block."""
if model_family == 'poisson_gamma':
if node_type == 'user':
return np.concatenate([params['a'][idx], params['b'][idx]])
else:
return np.concatenate([params['c'][idx], params['d'][idx]])
elif model_family == 'gaussian_gaussian':
if node_type == 'user':
return np.concatenate([params['m_U'][idx], params['s_U'][idx]])
else:
return np.concatenate([params['m_V'][idx], params['s_V'][idx]])
elif model_family == 'gaussian_gamma_map':
if node_type == 'user':
return params['alpha'][idx].copy()
else:
return params['beta'][idx].copy()
else:
raise ValueError(f"Unknown model family: {model_family}")
def compute_all_param_vector(params: dict, model_family: str) -> np.ndarray:
"""Flatten all parameters into a single vector."""
if model_family == 'poisson_gamma':
return np.concatenate([params['a'].ravel(), params['b'].ravel(),
params['c'].ravel(), params['d'].ravel()])
elif model_family == 'gaussian_gaussian':
return np.concatenate([params['m_U'].ravel(), params['s_U'].ravel(),
params['m_V'].ravel(), params['s_V'].ravel()])
elif model_family == 'gaussian_gamma_map':
return np.concatenate([params['alpha'].ravel(), params['beta'].ravel()])
else:
raise ValueError(f"Unknown model family: {model_family}")
def compute_deletion_influence_by_distance(full_params, exact_params, edge_to_remove,
edges, N, M, model_family, max_radius=6):
"""Compute deletion influence Delta_u(z) = ||lambda_u^* - lambda_u^{\\z}||
grouped by graph distance from seed set."""
user_to_items, item_to_users, _ = build_adjacency(edges, N, M)
distances = get_deletion_neighborhood(edge_to_remove, user_to_items, item_to_users,
N, M, max_radius)
blocks_by_dist = get_blocks_at_distance(distances, N)
influence_by_dist = {}
for dist, blocks in blocks_by_dist.items():
influences = []
for node_type, idx in blocks:
v_full = compute_block_param_vector(full_params, node_type, idx, model_family)
v_exact = compute_block_param_vector(exact_params, node_type, idx, model_family)
delta = np.linalg.norm(v_full - v_exact)
influences.append(delta)
influence_by_dist[dist] = {
'mean': float(np.mean(influences)),
'std': float(np.std(influences)),
'median': float(np.median(influences)),
'max': float(np.max(influences)),
'n_blocks': len(influences),
'values': [float(v) for v in influences],
}
return influence_by_dist
def fit_exponential_decay(influence_by_dist: dict, min_shells: int = 3, eps: float = 1e-12):
"""Fit log(Delta(r)) = alpha - mu * r."""
distances = sorted(influence_by_dist.keys())
r_vals = []
log_vals = []
for r in distances:
mean_inf = influence_by_dist[r]['mean']
if mean_inf > eps and influence_by_dist[r]['n_blocks'] >= 2:
r_vals.append(r)
log_vals.append(np.log(mean_inf + eps))
if len(r_vals) < min_shells:
return {
'mu_emp': None,
'intercept': None,
'r_squared': None,
'n_shells': len(r_vals),
'valid': False,
}
r_arr = np.array(r_vals, dtype=float)
log_arr = np.array(log_vals)
slope, intercept, r_value, p_value, std_err = stats.linregress(r_arr, log_arr)
return {
'mu_emp': float(-slope),
'intercept': float(intercept),
'r_squared': float(r_value ** 2),
'n_shells': len(r_vals),
'p_value': float(p_value),
'std_err': float(std_err),
'valid': True,
}
# ============================================================
# Local approximation error
# ============================================================
def compute_local_error(local_params, exact_params, model_family):
"""Err_R(z) = ||lambda^(R)_local - lambda^{\\z}||"""
v_local = compute_all_param_vector(local_params, model_family)
v_exact = compute_all_param_vector(exact_params, model_family)
err = np.linalg.norm(v_local - v_exact)
rel_err = err / (1.0 + np.linalg.norm(v_exact))
return {
'param_error': float(err),
'relative_error': float(rel_err),
}
# ============================================================
# ELBO / objective gap
# ============================================================
def compute_objective_gap(model, edges_without, exact_params, approx_params):
"""Gap_R = L_{\\z}(lambda^{\\z}) - L_{\\z}(lambda^(R))."""
try:
obj_exact = model.compute_elbo(edges_without, exact_params) if hasattr(model, 'compute_elbo') \
else model.compute_objective(edges_without, exact_params)
obj_approx = model.compute_elbo(edges_without, approx_params) if hasattr(model, 'compute_elbo') \
else model.compute_objective(edges_without, approx_params)
return float(obj_exact - obj_approx)
except Exception as e:
return None
# ============================================================
# Weighted interaction statistics (chi)
# ============================================================
def compute_chi_poisson_gamma(edge_to_remove, edges, params, N, M, K,
a0=0.3, b0=1.0, c0=0.3, d0=1.0):
"""Compute chi statistics for Poisson-Gamma model."""
a, b, c, d = params['a'], params['b'], params['c'], params['d']
# Compute constants
a_min = np.min(a[a > 0]) if np.any(a > 0) else a0
b_min = np.min(b[b > 0]) if np.any(b > 0) else b0
c_min = np.min(c[c > 0]) if np.any(c > 0) else c0
d_min = np.min(d[d > 0]) if np.any(d > 0) else d0
a_max = np.max(a)
c_max = np.max(c)
C_x = 0.5 * polygamma(1, max(c_min, 1e-3)) + 0.5 / max(d_min, 1e-6)
C_0 = 1.0 / max(d_min, 1e-6) + c_max / max(d_min**2, 1e-12)
C_tilde_x = 0.5 * polygamma(1, max(a_min, 1e-3)) + 0.5 / max(b_min, 1e-6)
C_tilde_0 = 1.0 / max(b_min, 1e-6) + a_max / max(b_min**2, 1e-12)
# Build adjacency
user_to_items = defaultdict(list)
item_to_users = defaultdict(list)
edge_dict = {}
for i, j, x in edges:
user_to_items[i].append(j)
item_to_users[j].append(i)
edge_dict[(i, j)] = x
i_del, j_del, x_del = edge_to_remove
# chi_i
chi_i = sum(C_x * edge_dict.get((i_del, j), 0) + C_0 for j in user_to_items.get(i_del, []))
# chi_tilde_j
chi_tilde_j = sum(C_tilde_x * edge_dict.get((i, j_del), 0) + C_tilde_0
for i in item_to_users.get(j_del, []))
chi_max = max(chi_i, chi_tilde_j)
chi_sum = chi_i + chi_tilde_j
# Empirical alternatives
seed_degree = len(user_to_items.get(i_del, [])) + len(item_to_users.get(j_del, []))
seed_count_sum = sum(edge_dict.get((i_del, j), 0) for j in user_to_items.get(i_del, [])) + \
sum(edge_dict.get((i, j_del), 0) for i in item_to_users.get(j_del, []))
return {
'chi_i': float(chi_i),
'chi_tilde_j': float(chi_tilde_j),
'chi_max': float(chi_max),
'chi_sum': float(chi_sum),
'seed_degree': int(seed_degree),
'seed_count_sum': float(seed_count_sum),
'C_x': float(C_x),
'C_0': float(C_0),
'C_tilde_x': float(C_tilde_x),
'C_tilde_0': float(C_tilde_0),
}
def compute_chi_gaussian(edge_to_remove, edges, params, N, M, K, sigma_x,
model_family='gaussian_gaussian'):
"""Compute interaction proxy for Gaussian models."""
user_to_items = defaultdict(list)
item_to_users = defaultdict(list)
edge_dict = {}
for i, j, x in edges:
user_to_items[i].append(j)
item_to_users[j].append(i)
edge_dict[(i, j)] = x
i_del, j_del, x_del = edge_to_remove
prec_x = 1.0 / (sigma_x ** 2)
if model_family == 'gaussian_gaussian':
m_U, s_U = params['m_U'], params['s_U']
m_V, s_V = params['m_V'], params['s_V']
chi_i = sum(prec_x * np.sum(m_V[j]**2 + s_V[j]) for j in user_to_items.get(i_del, []))
chi_tilde_j = sum(prec_x * np.sum(m_U[i]**2 + s_U[i]) for i in item_to_users.get(j_del, []))
elif model_family == 'gaussian_gamma_map':
from src.model import GaussianGammaMAP
sp = lambda x: np.log1p(np.exp(np.clip(x, -20, 20)))
U = sp(params['alpha'])
V = sp(params['beta'])
chi_i = sum(prec_x * np.sum(V[j]**2) for j in user_to_items.get(i_del, []))
chi_tilde_j = sum(prec_x * np.sum(U[i]**2) for i in item_to_users.get(j_del, []))
chi_max = max(chi_i, chi_tilde_j)
chi_sum = chi_i + chi_tilde_j
seed_degree = len(user_to_items.get(i_del, [])) + len(item_to_users.get(j_del, []))
return {
'chi_i': float(chi_i),
'chi_tilde_j': float(chi_tilde_j),
'chi_max': float(chi_max),
'chi_sum': float(chi_sum),
'seed_degree': int(seed_degree),
}
# ============================================================
# Gradient interference proxy
# ============================================================
def compute_gradient_interference(full_params, exact_params, local_params, model_family):
"""Compute gradient interference proxy.
g_del = lambda^* - lambda^{\\z}
g_ret = lambda^{\\z} - lambda^(R)_local
I(z) = |sum_u <g_del_u, g_ret_u>|
"""
v_full = compute_all_param_vector(full_params, model_family)
v_exact = compute_all_param_vector(exact_params, model_family)
v_local = compute_all_param_vector(local_params, model_family)
g_del = v_full - v_exact
g_ret = v_exact - v_local
raw_interference = float(np.abs(np.dot(g_del, g_ret)))
norm_del = np.linalg.norm(g_del)
norm_ret = np.linalg.norm(g_ret)
eps = 1e-12
cosine_interference = raw_interference / (norm_del * norm_ret + eps)
return {
'interference_raw': float(raw_interference),
'interference_cosine': float(cosine_interference),
'g_del_norm': float(norm_del),
'g_ret_norm': float(norm_ret),
}
# ============================================================
# Full metrics for one deletion
# ============================================================
def compute_all_metrics(full_params, exact_params, local_params_by_radius,
warm_start_params, one_step_params,
edge_to_remove, edges, N, M, K,
model_family, model=None, radii=[1, 2, 3, 4],
model_kwargs=None):
"""Compute all metrics for one deletion."""
results = {}
# Influence by distance
influence = compute_deletion_influence_by_distance(
full_params, exact_params, edge_to_remove, edges, N, M, model_family,
max_radius=max(radii) + 2)
results['influence_by_distance'] = {str(k): v['mean'] for k, v in influence.items()}
results['influence_by_distance_full'] = influence
# Decay fit
decay = fit_exponential_decay(influence)
results['empirical_decay_mu'] = decay['mu_emp']
results['empirical_decay_r2'] = decay['r_squared']
results['decay_valid'] = decay['valid']
# Chi statistics
if model_kwargs is None:
model_kwargs = {}
if model_family == 'poisson_gamma':
chi = compute_chi_poisson_gamma(
edge_to_remove, edges, full_params, N, M, K,
a0=model_kwargs.get('a0', 0.3), b0=model_kwargs.get('b0', 1.0),
c0=model_kwargs.get('c0', 0.3), d0=model_kwargs.get('d0', 1.0))
else:
chi = compute_chi_gaussian(
edge_to_remove, edges, full_params, N, M, K,
sigma_x=model_kwargs.get('sigma_x', 1.0), model_family=model_family)
results['chi_seed_max'] = chi['chi_max']
results['chi_seed_sum'] = chi['chi_sum']
results['seed_degree'] = chi['seed_degree']
# Local errors by radius
for R in radii:
if R in local_params_by_radius:
err = compute_local_error(local_params_by_radius[R], exact_params, model_family)
results[f'error_R{R}'] = err['param_error']
results[f'rel_error_R{R}'] = err['relative_error']
# Interference
interf = compute_gradient_interference(
full_params, exact_params, local_params_by_radius[R], model_family)
results[f'interference_raw_R{R}'] = interf['interference_raw']
results[f'interference_cosine_R{R}'] = interf['interference_cosine']
# Warm start error
if warm_start_params is not None:
ws_err = compute_local_error(warm_start_params, exact_params, model_family)
results['error_warm_start'] = ws_err['param_error']
results['rel_error_warm_start'] = ws_err['relative_error']
# One-step error
if one_step_params is not None:
os_err = compute_local_error(one_step_params, exact_params, model_family)
results['error_one_step'] = os_err['param_error']
results['rel_error_one_step'] = os_err['relative_error']
return results
# ============================================================
# Bootstrap confidence intervals
# ============================================================
def bootstrap_ci(values, n_boot=1000, ci=0.95, seed=42):
"""Compute bootstrap CI for mean."""
rng = np.random.RandomState(seed)
values = np.array(values)
values = values[np.isfinite(values)]
if len(values) < 3:
return {'mean': float(np.mean(values)) if len(values) > 0 else 0,
'ci_low': float(np.nan), 'ci_high': float(np.nan),
'se': float(np.nan), 'n': len(values)}
boot_means = np.array([
np.mean(rng.choice(values, len(values), replace=True))
for _ in range(n_boot)
])
alpha = (1 - ci) / 2
ci_low = float(np.percentile(boot_means, 100 * alpha))
ci_high = float(np.percentile(boot_means, 100 * (1 - alpha)))
return {
'mean': float(np.mean(values)),
'ci_low': ci_low,
'ci_high': ci_high,
'se': float(np.std(boot_means)),
'n': len(values),
}
def compute_bootstrap_summary(df, group_cols, metric_cols, n_boot=1000):
"""Compute bootstrap CIs for all metrics grouped by specified columns."""
import pandas as pd
rows = []
available_groups = [c for c in group_cols if c in df.columns]
if not available_groups:
groups = [('all', df)]
else:
groups = df.groupby(available_groups)
for grp_name, grp_df in groups:
row = {}
if isinstance(grp_name, tuple):
for col, val in zip(available_groups, grp_name):
row[col] = val
elif available_groups:
row[available_groups[0]] = grp_name
for col in metric_cols:
if col in grp_df.columns:
vals = grp_df[col].dropna().values
ci = bootstrap_ci(vals, n_boot=n_boot)
row[f'{col}_mean'] = ci['mean']
row[f'{col}_ci_low'] = ci['ci_low']
row[f'{col}_ci_high'] = ci['ci_high']
row[f'{col}_se'] = ci['se']
row[f'{col}_n'] = ci['n']
rows.append(row)
return pd.DataFrame(rows)
|