howto

Advanced XGBoost tuning in Python

August 22, 2016

XGBoost is an advanced gradient boosting tree library. XGboost is natively integrated into DSS virtual machine learning, meaning that you can train XGBoost models without writing any code or using any custom model.

In this Howto, we are going to cover advanced optimization techniques that can help you go even further with your XGboost models, by using custom Python recipes (or Jupyter notebooks).

We assume that you are already familiar with how to train a model using Python code (for example with scikit-learn).

Parameters

Most of these parameters are directly available when you create a XGBoost model using the visual machine learning component of DSS: you don't actually need to code for this part.

XGBoost has a large number of advanced parameters, which can all affect the quality and speed of your model.

max_depth : int
    Maximum tree depth for base learners.
learning_rate : float
    Boosting learning rate (xgb's "eta")
n_estimators : int
    Number of boosted trees to fit.
silent : boolean
    Whether to print messages while running boosting.
objective : string
    Specify the learning task and the corresponding learning objective.
nthread : int
    Number of parallel threads used to run xgboost.
gamma : float
    Minimum loss reduction required to make a further partition 
    on a leaf node of the tree.
min_child_weight : int
    Minimum sum of instance weight(hessian) needed in a child.
max_delta_step : int
    Maximum delta step we allow each tree's weight estimation to be.
subsample : float
    Subsample ratio of the training instance.
colsample_bytree : float
    Subsample ratio of columns when constructing each tree.
base_score:
    The initial prediction score of all instances, global bias.
seed : int
    Random number seed.
missing : float, optional
    Value in the data which needs to be present as a missing value. 
    If None, defaults to np.nan.

You have 2 ways to control overfitting in xgboost:

  • Control the model complexity with max_depth, min_child_weight and gamma.
  • Add randomness to make training robust to noise with subsample and colsample_bytree.

Using a Sparse matrix

Xgboost can take in input sparse matrix. That's very useful because when you have categorical variables with high cardinality, you can convert them into dummies matrix without being out of memory!

For this we use a python function:

from pandas.core.categorical import Categorical
from scipy.sparse import csr_matrix
import numpy as np

def sparse_dummies(categorical_values):
    categories = Categorical.from_array(categorical_values)
    N = len(categorical_values)
    row_numbers = np.arange(N, dtype=np.int)
    ones = np.ones((N,))
    return csr_matrix( (ones, (row_numbers, categories.codes)) )

sparse_dummies(df.VAR_0001)

This return a sparse matrix of 3 columns, one by value of VAR_0001:

    <145231x3 sparse matrix of type '<type 'numpy.float64'>'
        with 145231 stored elements in Compressed Sparse Row format>

You can concatenate this matrix with other dummies matrix with the scipy hstack function:

from scipy.sparse import hstack
cat1 = sparse_dummies(df.VAR_0001)
cat2 = sparse_dummies(df.VAR_0002)
hstack((cat1,cat2), format="csr")
    <145231x7 sparse matrix of type '<type 'numpy.float64'>'
        with 290462 stored elements in Compressed Sparse Row format>

Early stopping.

When creating a XGBoost model using the visual machine learning component of DSS, it already automatically uses early stopping (you don't actually need to code to benefit forom this).

A really cool feature is early stopping. As you are going to learn more and more trees, you will overfit your training dataset. Early stopping enables you to specify a validation dataset and the number of iterations after which the algorithm should stop if the score on your validation dataset didn't increase.

To use it, you can specify in the fit method of the classifier an evaluation set, an evaluation method and the early stopping round number:

clf = xgb.XGBClassifier(n_estimators=10000)
eval_set  = [(train,y_train), (valid,y_valid)]
clf.fit(train, y_train, eval_set=eval_set, 
        eval_metric="auc", early_stopping_rounds=30)

Here, we set explicitly the n_estimators to a very large number. In your job log you'll see the score increasing on the dataset you put in the eval_set list:

    Will train until validation_1 error hasn't decreased in 30 rounds.
    [0]    validation_0-auc:0.733451   validation_1-auc:0.698659
    [1]    validation_0-auc:0.776699   validation_1-auc:0.731099
    [2]    validation_0-auc:0.789156   validation_1-auc:0.740601
    [3]    validation_0-auc:0.792534   validation_1-auc:0.744378
    [4]    validation_0-auc:0.800747   validation_1-auc:0.748260
    [5]    validation_0-auc:0.805586   validation_1-auc:0.750209
    [6]    validation_0-auc:0.810889   validation_1-auc:0.752157
    [7]    validation_0-auc:0.812459   validation_1-auc:0.752554
    [8]    validation_0-auc:0.812928   validation_1-auc:0.752733
    [9]    validation_0-auc:0.813815   validation_1-auc:0.753650
    [10]   validation_0-auc:0.814547   validation_1-auc:0.753750
    ...
    ...
    ...
    [271]  validation_0-auc:0.897922   validation_1-auc:0.782187
    [272]  validation_0-auc:0.898150   validation_1-auc:0.782179
    [273]  validation_0-auc:0.898150   validation_1-auc:0.782179
    [274]  validation_0-auc:0.898439   validation_1-auc:0.782225
    [275]  validation_0-auc:0.898439   validation_1-auc:0.782225
    [276]  validation_0-auc:0.898591   validation_1-auc:0.782219
    Stopping. Best iteration:
    [246]  validation_0-auc:0.894087   validation_1-auc:0.782487

Note that you can define your own evaluation metric instead.

Viewing features importance.

You can get the features importance easily in clf.booster().get_fscore()where clf is your trained classifier.

For example, we can use this in a Jupyter notebook:

features = [ "your list of features ..." ]
mapFeat = dict(zip(["f"+str(i) for i in range(len(features))],features))
ts = pd.Series(clf.booster().get_fscore())
ts.index = ts.reset_index()['index'].map(mapFeat)
ts.order()[-15:].plot(kind="barh", title=("features importance"))

features importance

Using Hyperopt for grid searching

Fine-tuning your XGBoost can be done by exploring the space of parameters possibilities. For this task, you can use the hyperopt package. Hyperopt is a Python library for optimizing over awkward search spaces with real-valued, discrete, and conditional dimensions.

Here an exemple of python recipe to use it:

import dataiku
import pandas as pd, numpy as np
from dataiku import pandasutils as pdu
from sklearn.metrics import roc_auc_score
import xgboost as xgb
from hyperopt import hp, fmin, tpe, STATUS_OK, Trials

train = dataiku.Dataset("train").get_dataframe()
valid = dataiku.Dataset("valid").get_dataframe()

y_train = train.target
y_valid = valid.target

del train["target"]
del valid["target"]

def objective(space):

    clf = xgb.XGBClassifier(n_estimators = 10000, 
                            max_depth = space['max_depth'],
                            min_child_weight = space['min_child_weight'],
                            subsample = space['subsample'])

    eval_set  = [( train, y_train), ( valid, y_valid)]

    clf.fit(train[col_train], y_train,
            eval_set=eval_set, eval_metric="auc", 
            early_stopping_rounds=30)

    pred = clf.predict_proba(valid)[:,1]
    auc = roc_auc_score(y_valid, pred)
    print "SCORE:", auc

    return{'loss':1-auc, 'status': STATUS_OK }


space ={
        'max_depth': hp.quniform("x_max_depth", 5, 30, 1),
        'min_child_weight': hp.quniform ('x_min_child', 1, 10, 1),
        'subsample': hp.uniform ('x_subsample', 0.8, 1)
    }


trials = Trials()
best = fmin(fn=objective,
            space=space,
            algo=tpe.suggest,
            max_evals=100,
            trials=trials)

print best

After loading your datasets of training and validation, we define our objective function.

This function trains a model, evaluates it and returns the error on the validation set. We define the space we want to explore: here, we want to try values from 5 to 30 for max_depth, from 1 to 10 for min_child_weight and from 0.8 to 1 for subsample.

Hyperopt will minimise this error in a maximum of 100 experiments.

More documentation about Hyperopt is available .