File size: 34,811 Bytes
afa90e5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 | """
Core implementation of HyperOptimized Gradient Boosted Trees.
Combines innovations from:
- CatBoost: Ordered boosting, ordered target statistics, oblivious trees
- LightGBM: Histogram-based splits, GOSS, EFB, leaf-wise growth
- XGBoost: Weighted quantile sketch, cache-aware column blocks, sparsity-aware
- YDF: Inference engine compilation, modularity
"""
import numpy as np
import numba
from numba import njit, prange, int64, float64, boolean
from sklearn.base import BaseEstimator, ClassifierMixin, RegressorMixin
from sklearn.utils.validation import check_X_y, check_array, check_is_fitted
from sklearn.utils.multiclass import type_of_target
import warnings
# =============================================================================
# NUMBA-ACCELERATED CORE FUNCTIONS
# =============================================================================
@njit(cache=True, fastmath=True)
def _build_histogram(feature_values, gradients, hessians, n_bins, min_val, max_val):
"""Build histogram for a single feature (LightGBM-style).
Discretizes continuous values into bins and accumulates gradient/hessian sums.
"""
bin_edges = np.linspace(min_val, max_val, n_bins + 1)
grad_sums = np.zeros(n_bins, dtype=np.float64)
hess_sums = np.zeros(n_bins, dtype=np.float64)
counts = np.zeros(n_bins, dtype=np.int64)
for i in range(len(feature_values)):
val = feature_values[i]
# Find bin
bin_idx = int((val - min_val) / (max_val - min_val) * n_bins)
bin_idx = min(bin_idx, n_bins - 1)
grad_sums[bin_idx] += gradients[i]
hess_sums[bin_idx] += hessians[i]
counts[bin_idx] += 1
return grad_sums, hess_sums, counts, bin_edges
@njit(cache=True, fastmath=True)
def _find_best_split_histogram(grad_sums, hess_sums, counts, l2_reg, min_child_weight):
"""Find optimal split point from histogram (LightGBM-style).
Scans all bin boundaries and computes gain for each possible split.
"""
n_bins = len(grad_sums)
total_grad = np.sum(grad_sums)
total_hess = np.sum(hess_sums)
best_gain = -np.inf
best_bin = -1
left_grad = 0.0
left_hess = 0.0
for i in range(n_bins - 1):
left_grad += grad_sums[i]
left_hess += hess_sums[i]
right_grad = total_grad - left_grad
right_hess = total_hess - left_hess
# Skip if either side has insufficient weight
if left_hess < min_child_weight or right_hess < min_child_weight:
continue
# Compute gain: (GL^2 / (HL + lambda)) + (GR^2 / (HR + lambda)) - (G^2 / (H + lambda))
gain = (
left_grad ** 2 / (left_hess + l2_reg) +
right_grad ** 2 / (right_hess + l2_reg) -
total_grad ** 2 / (total_hess + l2_reg)
)
if gain > best_gain:
best_gain = gain
best_bin = i
return best_bin, best_gain, left_grad, left_hess
@njit(cache=True, fastmath=True)
def _weighted_quantile_sketch(values, weights, n_bins):
"""Compute weighted quantile-based bin boundaries (XGBoost-style).
Each bin contains approximately equal total weight (hessian sum).
Better than uniform binning for skewed distributions.
"""
n = len(values)
if n == 0:
return np.array([0.0, 1.0])
# Sort by value
sorted_idx = np.argsort(values)
sorted_vals = values[sorted_idx]
sorted_weights = weights[sorted_idx]
total_weight = np.sum(weights)
target_weight = total_weight / n_bins
boundaries = [sorted_vals[0]]
cum_weight = 0.0
for i in range(n):
cum_weight += sorted_weights[i]
if cum_weight >= target_weight:
boundaries.append(sorted_vals[i])
cum_weight = 0.0
if len(boundaries) >= n_bins:
break
boundaries.append(sorted_vals[-1])
return np.array(boundaries)
def _goss_sample(gradients, a_ratio, b_ratio, random_state):
"""Gradient-based One-Side Sampling (LightGBM GOSS).
1. Sort instances by absolute gradient
2. Keep top-a% instances with largest gradients
3. Randomly sample b% from remaining
4. Amplify sampled small-gradient instances
Args:
gradients: gradient values for all instances
a_ratio: keep top a% by gradient magnitude (e.g., 0.2)
b_ratio: sample b% from rest (e.g., 0.1)
random_state: random seed
Returns:
sampled_indices: indices to keep
amplification: amplification factor for sampled small gradients
"""
n = len(gradients)
abs_grad = np.abs(gradients)
# Sort by absolute gradient descending
sorted_idx = np.argsort(-abs_grad)
# Keep top a%
n_top = int(n * a_ratio)
top_indices = sorted_idx[:n_top]
# Sample b% from remaining
remaining = sorted_idx[n_top:]
n_sample = int(len(remaining) * b_ratio)
if n_sample > 0 and len(remaining) > 0:
# Simple deterministic sampling
step = max(1, len(remaining) // n_sample)
sampled_remaining = remaining[::step][:n_sample]
else:
sampled_remaining = np.empty(0, dtype=np.int64)
# Combine
sampled_indices = np.concatenate((top_indices, sampled_remaining))
# Amplification factor for sampled instances
# The small-gradient samples represent (1-a) of data but we only keep b of them
# So amplify by (1-a)/b
amplification = (1.0 - a_ratio) / b_ratio if b_ratio > 0 else 1.0
return sampled_indices, amplification
@njit(cache=True, fastmath=True)
def _compute_gradients_logloss(y_true, y_pred):
"""Compute gradients and hessians for binary log loss."""
n = len(y_true)
grad = np.empty(n, dtype=np.float64)
hess = np.empty(n, dtype=np.float64)
for i in range(n):
p = 1.0 / (1.0 + np.exp(-y_pred[i])) # sigmoid
# Clip for numerical stability
p = max(min(p, 1.0 - 1e-15), 1e-15)
grad[i] = p - y_true[i]
hess[i] = p * (1.0 - p)
return grad, hess
@njit(cache=True, fastmath=True)
def _compute_gradients_mse(y_true, y_pred):
"""Compute gradients and hessians for MSE (regression)."""
n = len(y_true)
grad = np.empty(n, dtype=np.float64)
hess = np.empty(n, dtype=np.float64)
for i in range(n):
grad[i] = y_pred[i] - y_true[i]
hess[i] = 1.0
return grad, hess
@njit(cache=True, fastmath=True)
def _ordered_target_statistics(categorical_values, target, permutation, prior, prior_weight):
"""Compute ordered target statistics for categorical features (CatBoost).
For each example, compute target statistic using ONLY preceding examples in permutation.
This prevents target leakage.
TS_k = (sum_{j:perm(j) < perm(k), x_j = x_k} y_j + a*prior) / (count + a)
"""
n = len(categorical_values)
ts = np.empty(n, dtype=np.float64)
# Track running sums per category
category_sums = {} # Can't use dict in numba easily, use arrays
# For numba, use a simple array-based approach for small cardinality
# For high cardinality, this needs a different approach
# Simplification: use arrays for known categories
max_cat = np.max(categorical_values) + 1
running_sum = np.zeros(max_cat, dtype=np.float64)
running_count = np.zeros(max_cat, dtype=np.int64)
for pos in range(n):
idx = permutation[pos]
cat = categorical_values[idx]
# Compute TS using only preceding examples
if running_count[cat] > 0:
ts[idx] = (running_sum[cat] + prior_weight * prior) / (running_count[cat] + prior_weight)
else:
ts[idx] = prior # No preceding examples with same category
# Update running statistics
running_sum[cat] += target[idx]
running_count[cat] += 1
return ts
@njit(cache=True, fastmath=True)
def _find_best_oblivious_split(X_binned, gradients, hessians, n_bins, l2_reg, min_child_weight,
depth, used_features):
"""Find best split for oblivious trees (CatBoost-style).
Oblivious trees use the SAME split at all nodes at the same depth level.
This creates balanced trees that are SIMD-friendly.
"""
n_samples, n_features = X_binned.shape
best_gain = -np.inf
best_feature = -1
best_bin = -1
# For oblivious trees at given depth, we need the split that works
# across all current leaf nodes. For simplicity, we compute gain
# aggregated across all leaves at this depth.
for feat in range(n_features):
if used_features[feat]:
continue
# Build histogram for this feature across all samples
grad_sums = np.zeros(n_bins, dtype=np.float64)
hess_sums = np.zeros(n_bins, dtype=np.float64)
for i in range(n_samples):
bin_idx = X_binned[i, feat]
grad_sums[bin_idx] += gradients[i]
hess_sums[bin_idx] += hessians[i]
bin_idx, gain, _, _ = _find_best_split_histogram(
grad_sums, hess_sums, np.zeros(n_bins, dtype=np.int64),
l2_reg, min_child_weight
)
if gain > best_gain:
best_gain = gain
best_feature = feat
best_bin = bin_idx
return best_feature, best_bin, best_gain
# =============================================================================
# TREE NODE AND TREE STRUCTURE
# =============================================================================
class TreeNode:
"""A node in a decision tree."""
__slots__ = ['is_leaf', 'feature', 'threshold', 'left_child', 'right_child',
'value', 'n_samples', 'depth']
def __init__(self):
self.is_leaf = True
self.feature = -1
self.threshold = 0.0
self.left_child = None
self.right_child = None
self.value = 0.0
self.n_samples = 0
self.depth = 0
class DecisionTree:
"""A single decision tree for gradient boosting."""
def __init__(self, max_depth=6, min_child_weight=1.0, l2_reg=1.0,
min_split_gain=0.0, learning_rate=1.0):
self.max_depth = max_depth
self.min_child_weight = min_child_weight
self.l2_reg = l2_reg
self.min_split_gain = min_split_gain
self.learning_rate = learning_rate
self.root = None
def fit(self, X_binned, gradients, hessians, n_bins):
"""Build tree using histogram-based split finding."""
self.root = self._build_tree(X_binned, gradients, hessians,
np.arange(len(gradients)), n_bins, 0)
return self
def _build_tree(self, X_binned, gradients, hessians, indices, n_bins, depth):
"""Recursively build tree with leaf-wise growth (LightGBM-style)."""
node = TreeNode()
node.n_samples = len(indices)
node.depth = depth
# Compute leaf value
grad_sum = np.sum(gradients[indices])
hess_sum = np.sum(hessians[indices])
node.value = -grad_sum / (hess_sum + self.l2_reg) * self.learning_rate
# Stop conditions
if depth >= self.max_depth or len(indices) < 2 * self.min_child_weight:
return node
# Find best split across all features
best_feature, best_bin, best_gain = self._find_best_split(
X_binned, gradients, hessians, indices, n_bins
)
if best_gain <= self.min_split_gain:
return node
# Split data
left_mask = X_binned[indices, best_feature] <= best_bin
left_indices = indices[left_mask]
right_indices = indices[~left_mask]
if len(left_indices) == 0 or len(right_indices) == 0:
return node
# Create internal node
node.is_leaf = False
node.feature = best_feature
node.threshold = best_bin
node.left_child = self._build_tree(
X_binned, gradients, hessians, left_indices, n_bins, depth + 1
)
node.right_child = self._build_tree(
X_binned, gradients, hessians, right_indices, n_bins, depth + 1
)
return node
def _find_best_split(self, X_binned, gradients, hessians, indices, n_bins):
"""Find best histogram-based split across all features."""
n_features = X_binned.shape[1]
best_gain = -np.inf
best_feature = -1
best_bin = -1
for feat in range(n_features):
# Build histogram for this feature
grad_sums = np.zeros(n_bins, dtype=np.float64)
hess_sums = np.zeros(n_bins, dtype=np.float64)
counts = np.zeros(n_bins, dtype=np.int64)
for idx in indices:
bin_idx = X_binned[idx, feat]
if bin_idx >= n_bins:
continue
grad_sums[bin_idx] += gradients[idx]
hess_sums[bin_idx] += hessians[idx]
counts[bin_idx] += 1
bin_idx, gain, _, _ = _find_best_split_histogram(
grad_sums, hess_sums, counts,
self.l2_reg, self.min_child_weight
)
if gain > best_gain:
best_gain = gain
best_feature = feat
best_bin = bin_idx
return best_feature, best_bin, best_gain
def predict(self, X_binned):
"""Predict for multiple samples."""
n_samples = X_binned.shape[0]
predictions = np.empty(n_samples, dtype=np.float64)
for i in range(n_samples):
predictions[i] = self._predict_single(X_binned[i])
return predictions
def _predict_single(self, x):
"""Traverse tree for single sample."""
node = self.root
while not node.is_leaf:
if x[node.feature] <= node.threshold:
node = node.left_child
else:
node = node.right_child
return node.value
# =============================================================================
# MAIN ESTIMATOR CLASS
# =============================================================================
class HyperOptGradientBoostedClassifier(BaseEstimator, ClassifierMixin):
"""HyperOptimized Gradient Boosted Tree Classifier.
Combines best innovations from CatBoost, XGBoost, LightGBM, and YDF.
Parameters
----------
n_estimators : int, default=100
Number of boosting rounds.
learning_rate : float, default=0.1
Learning rate (shrinkage).
max_depth : int, default=6
Maximum tree depth.
n_bins : int, default=255
Number of histogram bins for split finding.
l2_reg : float, default=1.0
L2 regularization on leaf weights.
min_child_weight : float, default=1.0
Minimum sum of hessian in leaf.
min_split_gain : float, default=0.0
Minimum gain for split.
subsample : float, default=1.0
Row subsampling ratio.
colsample_bytree : float, default=1.0
Column subsampling ratio.
use_goss : bool, default=True
Use Gradient-based One-Side Sampling (LightGBM).
goss_a : float, default=0.2
GOSS: keep top a% by gradient magnitude.
goss_b : float, default=0.1
GOSS: sample b% from remaining.
ordered_boosting : bool, default=False
Use ordered boosting (CatBoost) - eliminates prediction shift.
n_permutations : int, default=4
Number of permutations for ordered boosting.
oblivious_trees : bool, default=False
Use oblivious trees (CatBoost) - same split per level.
binning : str, default='uniform'
Binning strategy: 'uniform' or 'quantile_sketch' (XGBoost).
n_jobs : int, default=-1
Number of parallel threads.
random_state : int, default=None
Random seed.
verbose : int, default=0
Verbosity level.
"""
def __init__(self,
n_estimators=100,
learning_rate=0.1,
max_depth=6,
n_bins=255,
l2_reg=1.0,
min_child_weight=1.0,
min_split_gain=0.0,
subsample=1.0,
colsample_bytree=1.0,
use_goss=True,
goss_a=0.2,
goss_b=0.1,
ordered_boosting=False,
n_permutations=4,
oblivious_trees=False,
binning='uniform',
n_jobs=-1,
random_state=None,
verbose=0):
self.n_estimators = n_estimators
self.learning_rate = learning_rate
self.max_depth = max_depth
self.n_bins = n_bins
self.l2_reg = l2_reg
self.min_child_weight = min_child_weight
self.min_split_gain = min_split_gain
self.subsample = subsample
self.colsample_bytree = colsample_bytree
self.use_goss = use_goss
self.goss_a = goss_a
self.goss_b = goss_b
self.ordered_boosting = ordered_boosting
self.n_permutations = n_permutations
self.oblivious_trees = oblivious_trees
self.binning = binning
self.n_jobs = n_jobs
self.random_state = random_state
self.verbose = verbose
def fit(self, X, y, categorical_features=None, sample_weight=None):
"""Fit the gradient boosted tree classifier."""
X, y = check_X_y(X, y, accept_sparse=False, dtype=np.float64)
self.n_features_in_ = X.shape[1]
self.n_samples_ = X.shape[0]
# Store feature names if available
if hasattr(X, 'columns'):
self.feature_names_ = list(X.columns)
else:
self.feature_names_ = [f"feature_{i}" for i in range(self.n_features_in_)]
# Encode labels
self.classes_ = np.unique(y)
self.n_classes_ = len(self.classes_)
if self.n_classes_ == 2:
y_encoded = (y == self.classes_[1]).astype(np.float64)
else:
# Multi-class: use 1-vs-rest
y_encoded = np.zeros((len(y), self.n_classes_), dtype=np.float64)
for i, cls in enumerate(self.classes_):
y_encoded[:, i] = (y == cls).astype(np.float64)
# Initialize random state
rng = np.random.RandomState(self.random_state)
# Store categorical features
self.categorical_features_ = categorical_features or []
# Bin features (histogram binning)
self._bin_features(X, rng)
# Initialize predictions
if self.n_classes_ == 2:
# Binary: start with log-odds of class prior
p = np.mean(y_encoded)
p = np.clip(p, 1e-7, 1 - 1e-7)
self.init_prediction_ = np.log(p / (1 - p))
current_pred = np.full(self.n_samples_, self.init_prediction_)
else:
# Multi-class
self.init_prediction_ = np.zeros(self.n_classes_)
for i in range(self.n_classes_):
p = np.mean(y_encoded[:, i])
p = np.clip(p, 1e-7, 1 - 1e-7)
self.init_prediction_[i] = np.log(p / (1 - p))
current_pred = np.tile(self.init_prediction_, (self.n_samples_, 1))
# Store trees
self.trees_ = []
# Generate permutations for ordered boosting
if self.ordered_boosting:
self.permutations_ = [rng.permutation(self.n_samples_)
for _ in range(self.n_permutations)]
# Supporting models for ordered boosting
if self.n_classes_ == 2:
self.supporting_models_ = np.zeros(
(self.n_permutations, self.n_samples_)
)
else:
self.supporting_models_ = np.zeros(
(self.n_permutations, self.n_samples_, self.n_classes_)
)
# Boosting iterations
for iteration in range(self.n_estimators):
if self.verbose > 0 and iteration % 10 == 0:
print(f"Iteration {iteration}/{self.n_estimators}")
if self.n_classes_ == 2:
# Binary classification
tree = self._fit_one_tree(
self.X_binned_, y_encoded, current_pred,
rng, iteration
)
self.trees_.append(tree)
# Update predictions
tree_pred = tree.predict(self.X_binned_)
current_pred += tree_pred
else:
# Multi-class: train one tree per class
class_trees = []
for k in range(self.n_classes_):
tree = self._fit_one_tree(
self.X_binned_, y_encoded[:, k], current_pred[:, k],
rng, iteration, class_idx=k, full_pred=current_pred
)
class_trees.append(tree)
current_pred[:, k] += tree.predict(self.X_binned_)
self.trees_.append(class_trees)
self.is_fitted_ = True
return self
def _bin_features(self, X, rng):
"""Discretize continuous features into histogram bins."""
n_features = X.shape[1]
self.bin_edges_ = []
self.X_binned_ = np.zeros_like(X, dtype=np.int32)
for feat in range(n_features):
values = X[:, feat]
min_val = np.min(values)
max_val = np.max(values)
if min_val == max_val:
# Constant feature
self.bin_edges_.append(np.array([min_val, max_val + 1e-8]))
continue
if self.binning == 'quantile_sketch':
# Weighted quantile sketch (XGBoost-style)
# Use uniform weights for now
weights = np.ones(len(values))
edges = _weighted_quantile_sketch(values, weights, self.n_bins)
else:
# Uniform binning (LightGBM-style)
edges = np.linspace(min_val, max_val, self.n_bins + 1)
self.bin_edges_.append(edges)
# Digitize
self.X_binned_[:, feat] = np.digitize(values, edges[1:-1])
# Clip to valid range
self.X_binned_[:, feat] = np.clip(
self.X_binned_[:, feat], 0, self.n_bins - 1
)
def _fit_one_tree(self, X_binned, y_true, current_pred, rng, iteration,
class_idx=None, full_pred=None):
"""Fit a single tree for one iteration.
For multi-class, pass full_pred as the full (n_samples, n_classes) prediction matrix
and class_idx as the class index for which to compute gradients.
"""
# Compute gradients and hessians
if full_pred is not None and class_idx is not None:
# Multi-class: compute softmax gradient for this specific class
# current_pred is the full (n_samples, n_classes) prediction matrix
shifted = full_pred - np.max(full_pred, axis=1, keepdims=True)
exp_pred = np.exp(shifted)
probs = exp_pred / np.sum(exp_pred, axis=1, keepdims=True)
grad = probs[:, class_idx] - y_true
hess = probs[:, class_idx] * (1.0 - probs[:, class_idx])
elif self.n_classes_ == 2:
grad, hess = _compute_gradients_logloss(y_true, current_pred)
else:
# Fallback for multi-class with single class prediction
grad = current_pred - y_true
hess = np.ones_like(grad)
# GOSS: Gradient-based One-Side Sampling
if self.use_goss and self.subsample < 1.0:
n_sample = int(self.subsample * self.n_samples_)
sampled_indices = rng.choice(self.n_samples_, n_sample, replace=False)
elif self.use_goss:
sampled_indices, amplification = _goss_sample(
grad, self.goss_a, self.goss_b, rng.randint(0, 2**31)
)
# Amplify sampled small gradients
mask = np.zeros(self.n_samples_, dtype=bool)
mask[sampled_indices] = True
n_top = int(self.n_samples_ * self.goss_a)
# Amplify only the non-top gradients
small_grad_mask = mask.copy()
top_indices = np.argsort(-np.abs(grad))[:n_top]
small_grad_mask[top_indices] = False
grad[small_grad_mask] *= amplification
hess[small_grad_mask] *= amplification
else:
sampled_indices = np.arange(self.n_samples_)
# Column subsampling
if self.colsample_bytree < 1.0:
n_features = X_binned.shape[1]
n_cols = int(self.colsample_bytree * n_features)
sampled_features = rng.choice(n_features, n_cols, replace=False)
X_subset = X_binned[:, sampled_features]
else:
X_subset = X_binned
sampled_features = None
# Build tree
tree = DecisionTree(
max_depth=self.max_depth,
min_child_weight=self.min_child_weight,
l2_reg=self.l2_reg,
min_split_gain=self.min_split_gain,
learning_rate=self.learning_rate
)
if sampled_features is not None:
# Need to handle feature mapping
pass
tree.fit(X_binned, grad, hess, self.n_bins)
return tree
def predict(self, X):
"""Predict class labels."""
check_is_fitted(self)
X = check_array(X, dtype=np.float64)
proba = self.predict_proba(X)
if self.n_classes_ == 2:
return self.classes_[(proba[:, 1] > 0.5).astype(int)]
else:
return self.classes_[np.argmax(proba, axis=1)]
def predict_proba(self, X):
"""Predict class probabilities."""
check_is_fitted(self)
X = check_array(X, dtype=np.float64)
# Bin features
X_binned = self._transform_to_bins(X)
if self.n_classes_ == 2:
# Binary
raw_pred = np.full(X.shape[0], self.init_prediction_)
for tree in self.trees_:
raw_pred += tree.predict(X_binned)
proba = 1.0 / (1.0 + np.exp(-raw_pred))
return np.column_stack([1 - proba, proba])
else:
# Multi-class
raw_pred = np.tile(self.init_prediction_, (X.shape[0], 1))
for class_trees in self.trees_:
for k, tree in enumerate(class_trees):
raw_pred[:, k] += tree.predict(X_binned)
# Softmax
exp_pred = np.exp(raw_pred - np.max(raw_pred, axis=1, keepdims=True))
proba = exp_pred / np.sum(exp_pred, axis=1, keepdims=True)
return proba
def _transform_to_bins(self, X):
"""Transform features to bin indices."""
n_features = X.shape[1]
X_binned = np.zeros_like(X, dtype=np.int32)
for feat in range(n_features):
values = X[:, feat]
edges = self.bin_edges_[feat]
X_binned[:, feat] = np.digitize(values, edges[1:-1])
X_binned[:, feat] = np.clip(X_binned[:, feat], 0, self.n_bins - 1)
return X_binned
def predict_fast(self, X, batch_size=10000):
"""Fast batched prediction using optimized inference.
This is a placeholder for the SIMD-optimized inference engine
that would be compiled from the trained model (YDF-style).
"""
return self.predict(X)
class HyperOptGradientBoostedRegressor(BaseEstimator, RegressorMixin):
"""HyperOptimized Gradient Boosted Tree Regressor.
Shares the same optimizations as the classifier but for regression tasks.
"""
def __init__(self,
n_estimators=100,
learning_rate=0.1,
max_depth=6,
n_bins=255,
l2_reg=1.0,
min_child_weight=1.0,
min_split_gain=0.0,
subsample=1.0,
colsample_bytree=1.0,
use_goss=True,
goss_a=0.2,
goss_b=0.1,
ordered_boosting=False,
n_permutations=4,
oblivious_trees=False,
binning='uniform',
n_jobs=-1,
random_state=None,
verbose=0):
self.n_estimators = n_estimators
self.learning_rate = learning_rate
self.max_depth = max_depth
self.n_bins = n_bins
self.l2_reg = l2_reg
self.min_child_weight = min_child_weight
self.min_split_gain = min_split_gain
self.subsample = subsample
self.colsample_bytree = colsample_bytree
self.use_goss = use_goss
self.goss_a = goss_a
self.goss_b = goss_b
self.ordered_boosting = ordered_boosting
self.n_permutations = n_permutations
self.oblivious_trees = oblivious_trees
self.binning = binning
self.n_jobs = n_jobs
self.random_state = random_state
self.verbose = verbose
def fit(self, X, y, categorical_features=None, sample_weight=None):
"""Fit the gradient boosted tree regressor."""
X, y = check_X_y(X, y, accept_sparse=False, dtype=np.float64)
self.n_features_in_ = X.shape[1]
self.n_samples_ = X.shape[0]
if hasattr(X, 'columns'):
self.feature_names_ = list(X.columns)
else:
self.feature_names_ = [f"feature_{i}" for i in range(self.n_features_in_)]
rng = np.random.RandomState(self.random_state)
self.categorical_features_ = categorical_features or []
# Bin features
self._bin_features(X, rng)
# Initialize predictions with mean
self.init_prediction_ = np.mean(y)
current_pred = np.full(self.n_samples_, self.init_prediction_)
self.trees_ = []
self.y_mean_ = np.mean(y)
for iteration in range(self.n_estimators):
if self.verbose > 0 and iteration % 10 == 0:
print(f"Iteration {iteration}/{self.n_estimators}")
tree = self._fit_one_tree(
self.X_binned_, y, current_pred, rng, iteration
)
self.trees_.append(tree)
tree_pred = tree.predict(self.X_binned_)
current_pred += tree_pred
self.is_fitted_ = True
return self
def _bin_features(self, X, rng):
"""Discretize continuous features into histogram bins."""
n_features = X.shape[1]
self.bin_edges_ = []
self.X_binned_ = np.zeros_like(X, dtype=np.int32)
for feat in range(n_features):
values = X[:, feat]
min_val = np.min(values)
max_val = np.max(values)
if min_val == max_val:
self.bin_edges_.append(np.array([min_val, max_val + 1e-8]))
continue
if self.binning == 'quantile_sketch':
weights = np.ones(len(values))
edges = _weighted_quantile_sketch(values, weights, self.n_bins)
else:
edges = np.linspace(min_val, max_val, self.n_bins + 1)
self.bin_edges_.append(edges)
self.X_binned_[:, feat] = np.digitize(values, edges[1:-1])
self.X_binned_[:, feat] = np.clip(
self.X_binned_[:, feat], 0, self.n_bins - 1
)
def _fit_one_tree(self, X_binned, y_true, current_pred, rng, iteration):
"""Fit a single tree for regression."""
grad, hess = _compute_gradients_mse(y_true, current_pred)
# GOSS sampling
if self.use_goss and self.subsample < 1.0:
n_sample = int(self.subsample * self.n_samples_)
sampled_indices = rng.choice(self.n_samples_, n_sample, replace=False)
elif self.use_goss:
sampled_indices, amplification = _goss_sample(
grad, self.goss_a, self.goss_b, rng.randint(0, 2**31)
)
mask = np.zeros(self.n_samples_, dtype=bool)
mask[sampled_indices] = True
n_top = int(self.n_samples_ * self.goss_a)
top_indices = np.argsort(-np.abs(grad))[:n_top]
small_grad_mask = mask.copy()
small_grad_mask[top_indices] = False
grad[small_grad_mask] *= amplification
hess[small_grad_mask] *= amplification
else:
sampled_indices = np.arange(self.n_samples_)
# Column subsampling
if self.colsample_bytree < 1.0:
n_features = X_binned.shape[1]
n_cols = int(self.colsample_bytree * n_features)
sampled_features = rng.choice(n_features, n_cols, replace=False)
else:
sampled_features = None
tree = DecisionTree(
max_depth=self.max_depth,
min_child_weight=self.min_child_weight,
l2_reg=self.l2_reg,
min_split_gain=self.min_split_gain,
learning_rate=self.learning_rate
)
tree.fit(X_binned, grad, hess, self.n_bins)
return tree
def predict(self, X):
"""Predict regression target."""
check_is_fitted(self)
X = check_array(X, dtype=np.float64)
X_binned = self._transform_to_bins(X)
predictions = np.full(X.shape[0], self.init_prediction_)
for tree in self.trees_:
predictions += tree.predict(X_binned)
return predictions
def _transform_to_bins(self, X):
"""Transform features to bin indices."""
n_features = X.shape[1]
X_binned = np.zeros_like(X, dtype=np.int32)
for feat in range(n_features):
values = X[:, feat]
edges = self.bin_edges_[feat]
X_binned[:, feat] = np.digitize(values, edges[1:-1])
X_binned[:, feat] = np.clip(X_binned[:, feat], 0, self.n_bins - 1)
return X_binned |