If you like what we're working on, please  star us on GitHub. This enables us to continue to give back to the community.


Extreme Gradient Boosting (XGBoost) is a generalized gradient-boosted decision tree (GBDT) ML toolkit that is scalable. It is the top machine learning library for regression, classification, and ranking tasks, and it includes parallel tree boosting.

To understand XGBoost, you must first understand the machine learning ideas and methods on which it is based: supervised machine learning, decision trees, ensemble learning, and gradient boosting.

Supervised machine learning use algorithms to train a model to detect patterns in a dataset with labels and features, and then to predict labels on fresh dataset features using the learned model.

By analyzing a tree of if-then-else true/false feature questions and determining the least number of questions needed to assess the probability of reaching a good decision, decision trees produce a model that predicts the label. Decision trees can be used to forecast a continuous numeric value or for categorization to predict a category. A decision tree is used in the basic example below to estimate a property price depending on the size and number of bedrooms.

Gradient Boosting Decision Trees (GBDT) is a decision tree ensemble learning approach for classification and regression that is comparable to a random forest. To create a better model, ensemble learning techniques mix different machine learning methods.

Random forest and GBDT both create a model with many decision trees. The distinction is in the way the trees are constructed and joined. Random forest builds entire decision trees in parallel using random bootstrap samples of the data set using a method called bagging. The average of all decision tree forecasts is used to make the final prediction.

Gradient boosting refers to the process of “boosting” or enhancing a single weak model by merging it with several additional weak models to create a collectively strong model. In gradient boosting, which is an extension of boosting, the approach of additively building weak models is specified as a gradient descent algorithm. To reduce mistakes, gradient boosting specifies intended outcomes for the next model. The gradient of the error (thus the name gradient boosting) concerning the prediction determines the targeted outcomes for each case.

GBDTs train an assembly of thin decision trees repeatedly, with each iteration using the prior model’s error residuals to fit the next model. The weighted total of all tree forecasts yields the final prediction. GBDT “boosting” lowers bias and underfitting, whereas random forest “bagging” minimizes variance and overfitting.

  • XGBoost is a flexible and extremely accurate version of gradient boosting.

It was designed primarily to increase machine learning model performance and computational speed. Unlike GBDT, XGBoost builds trees in parallel rather than sequentially. It employs a level-wise technique, scanning over gradient values and evaluating the quality of splits at each feasible split in the training set using partial sums.

Optimization and enhancements 

  • Pruning: Within the GBM framework, the stopping criterion for tree splitting is greedy and is based on the negative loss criterion at the split point. XGBoost starts pruning trees backward, using the max depth argument instead of the criteria. This ‘depth-first’ method dramatically enhances computing performance.
  • Hardware Optimization: This method was created to make the most of the hardware resources available. This is performed by cache awareness, which involves each thread creating internal buffers to hold gradient statistics. Out-of-core computing, for example, optimizes available disk space while managing large data-frames that don’t fit in memory.
  • Regularization: To minimize overfitting, it penalizes more complicated models using both LASSO and Ridge regularization.
  • Weighted Quantile Sketch: To successfully locate the best split points among weighted datasets, XGBoost uses the distributed weighted Quantile Sketch technique.

XGBoost has acquired a lot of traction in recent years as a result of its ability to assist individuals and teams win almost every Kaggle structured data competition. Companies and researchers submit data to these contests, and data miners compete to create the best models for predicting and explaining the data.

Initially, XGBoost implementations in Python and R were created. XGBoost now includes package implementations for Scala, Java, and other languages as a result of its popularity. These enhancements have broadened the attractiveness of the XGBoost library to even more developers in the Kaggle community.

XGBoost is compatible with many different tools and packages, including scikit-learn for Python users and caret for R users. XGBoost is also compatible with distributed processing frameworks such as Apache Spark and Dask.

To find the champion algorithm, data scientists must evaluate all potential algorithms for the data at hand. Furthermore, selecting the appropriate algorithm is insufficient. They must also tune the hyper-parameters to get the best algorithm configuration for a dataset. There are also various additional factors to consider while selecting the winning method, such as computational complexity, explainability, and implementation simplicity.

Open source package for ml validation

Build Test Suites for ML Models & Data with Deepchecks

Get StartedOur GithubOur Github

Identifying and Preventing Key ML PitfallsDec 5th, 2022    06:00 PM PST

Register NowRegister Now