Notes on Parameter Tuning

Parameter tuning is a dark art in machine learning, the optimal parameters of a model can depend on many scenarios. So it is impossible to create a comprehensive guide for doing so.

This document tries to provide some guideline for parameters in XGBoost.

Understanding Bias-Variance Tradeoff

If you take a machine learning or statistics course, this is likely to be one of the most important concepts. When we allow the model to get more complicated (e.g. more depth), the model has better ability to fit the training data, resulting in a less biased model. However, such complicated model requires more data to fit.

Most of parameters in XGBoost are about bias variance tradeoff. The best model should trade the model complexity with its predictive power carefully. Parameters Documentation will tell you whether each parameter will make the model more conservative or not. This can be used to help you turn the knob between complicated model and simple model.

Control Overfitting

When you observe high training accuracy, but low test accuracy, it is likely that you encountered overfitting problem.

There are in general two ways that you can control overfitting in XGBoost:

  • The first way is to directly control model complexity.

    • This includes max_depth, min_child_weight, gamma, max_cat_threshold and other similar regularization parameters. See XGBoost Parameters for a comprehensive set of parameters.

    • Set a constant base_score based on your own criteria. See Intercept for more info.

  • The second way is to add randomness to make training robust to noise.

    • This includes subsample and colsample_bytree, which may be used with boosting RF num_parallel_tree.

    • You can also reduce stepsize eta, possibly with a training callback. Remember to increase num_round when you do so.

Handle Imbalanced Dataset

For common cases such as ads clickthrough log, the dataset is extremely imbalanced. This can affect the training of XGBoost model, and there are two ways to improve it.

  • If you care only about the overall performance metric (AUC) of your prediction

    • Balance the positive and negative weights via scale_pos_weight

    • Use AUC for evaluation

  • If you care about predicting the right probability

    • In such a case, you cannot re-balance the dataset

    • Set parameter max_delta_step to a finite number (say 1) to help convergence

Use Hyper Parameter Optimization (HPO) Frameworks

Tuning models is a sophisticated task and there are advanced frameworks to help you. For examples, some meta estimators in scikit-learn like sklearn.model_selection.HalvingGridSearchCV can help guide the search process. Optuna is another great option and there are many more based on different branches of statistics.

Know Your Data

It cannot be stressed enough the importance of understanding the data, sometimes that’s all it takes to get a good model. Many solutions use a simple XGBoost tree model without much tuning and emphasize the data pre-processing step. XGBoost can help feature selection by providing both a global feature importance score and sample feature importance with SHAP value. Also, there are parameters specifically targeting categorical features, and tasks like survival and ranking. Feel free to explore them.

Reducing Memory Usage

If you are using a HPO library like sklearn.model_selection.GridSearchCV, please control the number of threads it can use. It’s best to let XGBoost to run in parallel instead of asking GridSearchCV to run multiple experiments at the same time. For instance, creating a fold of data for cross validation can consume a significant amount of memory:

# This creates a copy of dataset. X and X_train are both in memory at the same time.

# This happens for every thread at the same time if you run `GridSearchCV` with
# `n_jobs` larger than 1

X_train, X_test, y_train, y_test = train_test_split(X, y)

df = pd.DataFrame()
# This creates a new copy of the dataframe, even if you specify the inplace parameter
new_df = df.drop(...)

array = np.array(...)
# This may or may not make a copy of the data, depending on the type of the data
array.astype(np.float32)

# np by default uses double, do you actually need it?
array = np.array(...)

You can find some more specific memory reduction practices scattered through the documents For instances: Distributed XGBoost with Dask, XGBoost GPU Support. However, before going into these, being conscious about making data copies is a good starting point. It usually consumes a lot more memory than people expect.