XGBoost from Scratch
Want to share your content on python-bloggers? click here.
Well, dear reader, it’s that time again, time for us to do a seemingly unnecessary scratch build of a popular algorithm that most people would simply import from the library without a second thought. But readers of this blog are not most people. Of course you know that when we do scratch builds, it’s not for the hell of it, it’s for the purpose of demystification. To that end, today we are going to implement XGBoost from scratch in python, using only numpy and pandas.
Specifically we’re going to implement the core statistical learning algorithm of XGBoost, including most of the key hyperparameters and their functionality. Our implementation will also support user-defined custom objective functions, meaning that it can perform regression, classification, and whatever exotic learning tasks you can dream up, as long as you can write down a twice-differentiable objective function. We’ll refrain from implementing some simple features like column subsampling which will be left to you, gentle reader, as exercises. In terms of tree methods, we’re going to implement the exact tree-splitting algorithm, leaving the sparsity-aware method (used to handle missing feature values) and the approximate method (used for scalability) as exercises or maybe topics for future posts.
As always, if something is unclear, try backtracking through the previous posts on gradient boosting and decision trees to clarify your intuition. We’ve already built up all the statistical and computational background needed to make sense of this scratch build. Here are the most important prerequisite posts:
Great, let’s do this.
The XGBoost Model Class
We begin with the user-facing API for our model, a class called XGBoostModel which will implement gradient boosting and prediction. To be more consistent with the XGBoost library, we’ll pass hyperparameters to our model in a parameter dictionary, so our init method is going to pull relevant parameters out of the dictionary and set them as object attributes. Note the use of python’s defaultdict so we don’t have to worry about handling key errors if we try to access a parameter that the user didn’t set in the dictionary.
import math import numpy as np import pandas as pd from collections import defaultdict
class XGBoostModel():
'''XGBoost from Scratch
'''
def __init__(self, params, random_seed=None):
self.params = defaultdict(lambda: None, params)
self.subsample = self.params['subsample'] \
if self.params['subsample'] else 1.0
self.learning_rate = self.params['learning_rate'] \
if self.params['learning_rate'] else 0.3
self.base_prediction = self.params['base_score'] \
if self.params['base_score'] else 0.5
self.max_depth = self.params['max_depth'] \
if self.params['max_depth'] else 5
self.rng = np.random.default_rng(seed=random_seed)The fit method, based on our classic GBM, takes a feature dataframe, a target vector, the objective function, and the number of boosting rounds as arguments. The user-supplied objective function should be an object with loss, gradient, and hessian methods, each of which takes a target vector and a prediction vector as input; the loss method should return a scalar loss score, the gradient method should return a vector of gradients, and the hessian method should return a vector of hessians.
In contrast to boosting in the classic GBM, instead of computing residuals between the current predictions and the target, we compute gradients and hessians of the loss function with respect to the current predictions, and instead of predicting residuals with a decision tree, we fit a special XGBoost tree booster (which we’ll implement in a moment) using the gradients and hessians. I’ve also added row subsampling by drawing a random subset of instance indices and passing them to the tree booster during each boosting round. The rest of the fit method is the same as the classic GBM, and the predict method is identical too.
def fit(self, X, y, objective, num_boost_round, verbose=False):
current_predictions = self.base_prediction * np.ones(shape=y.shape)
self.boosters = []
for i in range(num_boost_round):
gradients = objective.gradient(y, current_predictions)
hessians = objective.hessian(y, current_predictions)
sample_idxs = None if self.subsample == 1.0 \
else self.rng.choice(len(y),
size=math.floor(self.subsample*len(y)),
replace=False)
booster = TreeBooster(X, gradients, hessians,
self.params, self.max_depth, sample_idxs)
current_predictions += self.learning_rate * booster.predict(X)
self.boosters.append(booster)
if verbose:
print(f'[{i}] train loss = {objective.loss(y, current_predictions)}')
def predict(self, X):
return (self.base_prediction + self.learning_rate
* np.sum([booster.predict(X) for booster in self.boosters], axis=0))
XGBoostModel.fit = fit
XGBoostModel.predict = predict All we have to do now is implement the tree booster.
The XGBoost Tree Booster
The XGBoost tree booster is a modified version of the decision tree that we built in the decision tree from scratch post. Like the decision tree, we recursively build a binary tree structure by finding the best split rule for each node in the tree. The main difference is the criterion for evaluating splits and the way that we define a leaf’s predicted value. Instead of being functions of the target values of the instances in each node, the criterion and predicted values are functions of the instance gradients and hessians. Thus we need only make a couple of modifications to our previous decision tree implementation to create the XGBoost tree booster.
Initialization and Inserting Child Nodes
Most of the init method is just parsing the parameter dictionary to assign parameters as object attributes. The one notable difference from our decision tree is in the way we define the node’s predicted value. We define self.value according to equation 5 of the XGBoost paper, a simple function of the gradient and hessian values of the instances in the current node. Of course the init also goes on to build the tree via the maybe insert child nodes method. This method is nearly identical to the one we implemented for our decision tree. So far so good.
class TreeBooster():
def __init__(self, X, g, h, params, max_depth, idxs=None):
self.params = params
self.max_depth = max_depth
assert self.max_depth >= 0, 'max_depth must be nonnegative'
self.min_child_weight = params['min_child_weight'] \
if params['min_child_weight'] else 1.0
self.reg_lambda = params['reg_lambda'] if params['reg_lambda'] else 1.0
self.gamma = params['gamma'] if params['gamma'] else 0.0
self.colsample_bynode = params['colsample_bynode'] \
if params['colsample_bynode'] else 1.0
if isinstance(g, pd.Series): g = g.values
if isinstance(h, pd.Series): h = h.values
if idxs is None: idxs = np.arange(len(g))
self.X, self.g, self.h, self.idxs = X, g, h, idxs
self.n, self.c = len(idxs), X.shape[1]
self.value = -g[idxs].sum() / (h[idxs].sum() + self.reg_lambda) # Eq (5)
self.best_score_so_far = 0.
if self.max_depth > 0:
self._maybe_insert_child_nodes()
def _maybe_insert_child_nodes(self):
for i in range(self.c): self._find_better_split(i)
if self.is_leaf: return
x = self.X.values[self.idxs,self.split_feature_idx]
left_idx = np.nonzero(x <= self.threshold)[0]
right_idx = np.nonzero(x > self.threshold)[0]
self.left = TreeBooster(self.X, self.g, self.h, self.params,
self.max_depth - 1, self.idxs[left_idx])
self.right = TreeBooster(self.X, self.g, self.h, self.params,
self.max_depth - 1, self.idxs[right_idx])
@property
def is_leaf(self): return self.best_score_so_far == 0.
def _find_better_split(self, feature_idx):
passSplit Finding
Split finding follows the exact same pattern that we used in the decision tree, except we keep track of gradient and hessian stats instead of target value stats, and of course we use the XGBoost gain criterion (equation 7 from the paper) for evaluating splits.
def _find_better_split(self, feature_idx):
x = self.X.values[self.idxs, feature_idx]
g, h = self.g[self.idxs], self.h[self.idxs]
sort_idx = np.argsort(x)
sort_g, sort_h, sort_x = g[sort_idx], h[sort_idx], x[sort_idx]
sum_g, sum_h = g.sum(), h.sum()
sum_g_right, sum_h_right = sum_g, sum_h
sum_g_left, sum_h_left = 0., 0.
for i in range(0, self.n - 1):
g_i, h_i, x_i, x_i_next = sort_g[i], sort_h[i], sort_x[i], sort_x[i + 1]
sum_g_left += g_i; sum_g_right -= g_i
sum_h_left += h_i; sum_h_right -= h_i
if sum_h_left < self.min_child_weight or x_i == x_i_next:continue
if sum_h_right < self.min_child_weight: break
gain = 0.5 * ((sum_g_left**2 / (sum_h_left + self.reg_lambda))
+ (sum_g_right**2 / (sum_h_right + self.reg_lambda))
- (sum_g**2 / (sum_h + self.reg_lambda))
) - self.gamma/2 # Eq(7) in the xgboost paper
if gain > self.best_score_so_far:
self.split_feature_idx = feature_idx
self.best_score_so_far = gain
self.threshold = (x_i + x_i_next) / 2
TreeBooster._find_better_split = _find_better_splitPrediction
Prediction works exactly the same as in our decision tree, and the methods are nearly identical.
def predict(self, X):
return np.array([self._predict_row(row) for i, row in X.iterrows()])
def _predict_row(self, row):
if self.is_leaf:
return self.value
child = self.left if row[self.split_feature_idx] <= self.threshold \
else self.right
return child._predict_row(row)
TreeBooster.predict = predict
TreeBooster._predict_row = _predict_row The Complete XGBoost From Scratch Implementation
Here’s the entire implementation which produces a usable XGBoostModel class with fit and predict methods.
class XGBoostModel():
'''XGBoost from Scratch
'''
def __init__(self, params, random_seed=None):
self.params = defaultdict(lambda: None, params)
self.subsample = self.params['subsample'] \
if self.params['subsample'] else 1.0
self.learning_rate = self.params['learning_rate'] \
if self.params['learning_rate'] else 0.3
self.base_prediction = self.params['base_score'] \
if self.params['base_score'] else 0.5
self.max_depth = self.params['max_depth'] \
if self.params['max_depth'] else 5
self.rng = np.random.default_rng(seed=random_seed)
def fit(self, X, y, objective, num_boost_round, verbose=False):
current_predictions = self.base_prediction * np.ones(shape=y.shape)
self.boosters = []
for i in range(num_boost_round):
gradients = objective.gradient(y, current_predictions)
hessians = objective.hessian(y, current_predictions)
sample_idxs = None if self.subsample == 1.0 \
else self.rng.choice(len(y),
size=math.floor(self.subsample*len(y)),
replace=False)
booster = TreeBooster(X, gradients, hessians,
self.params, self.max_depth, sample_idxs)
current_predictions += self.learning_rate * booster.predict(X)
self.boosters.append(booster)
if verbose:
print(f'[{i}] train loss = {objective.loss(y, current_predictions)}')
def predict(self, X):
return (self.base_prediction + self.learning_rate
* np.sum([booster.predict(X) for booster in self.boosters], axis=0))
class TreeBooster():
def __init__(self, X, g, h, params, max_depth, idxs=None):
self.params = params
self.max_depth = max_depth
assert self.max_depth >= 0, 'max_depth must be nonnegative'
self.min_child_weight = params['min_child_weight'] \
if params['min_child_weight'] else 1.0
self.reg_lambda = params['reg_lambda'] if params['reg_lambda'] else 1.0
self.gamma = params['gamma'] if params['gamma'] else 0.0
self.colsample_bynode = params['colsample_bynode'] \
if params['colsample_bynode'] else 1.0
if isinstance(g, pd.Series): g = g.values
if isinstance(h, pd.Series): h = h.values
if idxs is None: idxs = np.arange(len(g))
self.X, self.g, self.h, self.idxs = X, g, h, idxs
self.n, self.c = len(idxs), X.shape[1]
self.value = -g[idxs].sum() / (h[idxs].sum() + self.reg_lambda) # Eq (5)
self.best_score_so_far = 0.
if self.max_depth > 0:
self._maybe_insert_child_nodes()
def _maybe_insert_child_nodes(self):
for i in range(self.c): self._find_better_split(i)
if self.is_leaf: return
x = self.X.values[self.idxs,self.split_feature_idx]
left_idx = np.nonzero(x <= self.threshold)[0]
right_idx = np.nonzero(x > self.threshold)[0]
self.left = TreeBooster(self.X, self.g, self.h, self.params,
self.max_depth - 1, self.idxs[left_idx])
self.right = TreeBooster(self.X, self.g, self.h, self.params,
self.max_depth - 1, self.idxs[right_idx])
@property
def is_leaf(self): return self.best_score_so_far == 0.
def _find_better_split(self, feature_idx):
x = self.X.values[self.idxs, feature_idx]
g, h = self.g[self.idxs], self.h[self.idxs]
sort_idx = np.argsort(x)
sort_g, sort_h, sort_x = g[sort_idx], h[sort_idx], x[sort_idx]
sum_g, sum_h = g.sum(), h.sum()
sum_g_right, sum_h_right = sum_g, sum_h
sum_g_left, sum_h_left = 0., 0.
for i in range(0, self.n - 1):
g_i, h_i, x_i, x_i_next = sort_g[i], sort_h[i], sort_x[i], sort_x[i + 1]
sum_g_left += g_i; sum_g_right -= g_i
sum_h_left += h_i; sum_h_right -= h_i
if sum_h_left < self.min_child_weight or x_i == x_i_next:continue
if sum_h_right < self.min_child_weight: break
gain = 0.5 * ((sum_g_left**2 / (sum_h_left + self.reg_lambda))
+ (sum_g_right**2 / (sum_h_right + self.reg_lambda))
- (sum_g**2 / (sum_h + self.reg_lambda))
) - self.gamma/2 # Eq(7) in the xgboost paper
if gain > self.best_score_so_far:
self.split_feature_idx = feature_idx
self.best_score_so_far = gain
self.threshold = (x_i + x_i_next) / 2
def predict(self, X):
return np.array([self._predict_row(row) for i, row in X.iterrows()])
def _predict_row(self, row):
if self.is_leaf:
return self.value
child = self.left if row[self.split_feature_idx] <= self.threshold \
else self.right
return child._predict_row(row)Testing
Let’s take this baby for a spin and benchmark its performance against the actual XGBoost library. We use the scikit learn California housing dataset for benchmarking.
from sklearn.datasets import fetch_california_housing
from sklearn.model_selection import train_test_split
X, y = fetch_california_housing(as_frame=True, return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3,
random_state=43)Let’s start with a nice friendly squared error objective function for training. We should probably have a future post all about how to define custom objective functions in XGBoost, but for now, here’s how I define squared error.
class SquaredErrorObjective():
def loss(self, y, pred): return np.mean((y - pred)**2)
def gradient(self, y, pred): return pred - y
def hessian(self, y, pred): return np.ones(len(y))Here I use a more or less arbitrary set of hyperparameters for training. Feel free to play around with tuning and trying other parameter combinations yourself.
import xgboost as xgb
params = {
'learning_rate': 0.1,
'max_depth': 5,
'subsample': 0.8,
'reg_lambda': 1.5,
'gamma': 0.0,
'min_child_weight': 25,
'base_score': 0.0,
'tree_method': 'exact',
}
num_boost_round = 50
# train the from-scratch XGBoost model
model_scratch = XGBoostModel(params, random_seed=42)
model_scratch.fit(X_train, y_train, SquaredErrorObjective(), num_boost_round)
# train the library XGBoost model
dtrain = xgb.DMatrix(X_train, label=y_train)
dtest = xgb.DMatrix(X_test, label=y_test)
model_xgb = xgb.train(params, dtrain, num_boost_round)Let’s check the models’ performance on the held out test data to benchmark our implementation.
pred_scratch = model_scratch.predict(X_test)
pred_xgb = model_xgb.predict(dtest)
print(f'scratch score: {SquaredErrorObjective().loss(y_test, pred_scratch)}')
print(f'xgboost score: {SquaredErrorObjective().loss(y_test, pred_xgb)}')scratch score: 0.2434125759558149 xgboost score: 0.24123239765807963
Well, look at that! Our scratch-built SGBoost is looking pretty consistent with the library. Go us!
Wrapping Up
I’d say this is a pretty good milestone for us here at Random Realizations. We’ve been hammering away at the various concepts around gradient boosting, leaving a trail of equations and scratch-built algos in our wake. Today we put all of that together to create a legit scratch build of XGBoost, something that would have been out of reach for me before we embarked on this journey together over a year ago. To anyone with the patience to read through this stuff, cheers to you! I hope you’re learning and enjoying this as much as I am.
Reader Exercises
If you want to take this a step further and deepen your understanding and coding abilities, let me recommend some exercises for you.
- Implement column subsampling. XGBoost itself provides column subsampling by tree, by level, and by node. Try implementing by tree first, then try adding by level or by node as well. These should be pretty straightforward to do.
- Implement sparsity aware split finding for missing feature values (Algorithm 2 in the XGBoost paper). This will be a little more involved, since you’ll need to refactor and modify several parts of the tree booster class.
Want to share your content on python-bloggers? click here.