In the previous post, we talk about a very popular Boosting algorithm - Gradient Boosting Decision Tree. The key of GBM is using Gradient Descent to optimize the loss function. But why Gradient Descent, not other numeric optimization method? Is it the fastest optimization method? Is there any problem with Gradient Descent?

If the target function is convex, then we won't need to worry about any of the above problem. Gradient Descent will lead us to the close solution in one iteration. However in real world, target function is rarely convex. In this case, Gradient Descent is no longer the fastest method, and it has some other problem, including:

• Step length is hard to choose. Small step converge slow, big step may lead to zigzag
• Converge can be slow when close to the optimum

All above issues with gradient descent is because it only consider first order of target function. In other words it tries to use linear function to approximate target function and find the direction that loss reduce fastest. Following this logic, if we use second order polynomial to approximate target function, we shall get better estimation of the direction. This leads to Newton Raphson method.

We define L as loss function, g is the gradient, and h is the second order derivative. Below is a second order Taylor expansion of the loss function.

We want to find \(\epsilon\) that minimize L, in other words g=0 and h>0. We take first order derivative of above approximation and get below:

This is the numeric optimization method used in XGBoost - Newton Raphson.

Compared with other boosting algorithm, XGBoost has a few innovations in following areas, including:

• Second order numeric optimization method
• Regularized model formulation
• Algo acceleration

Next let's dig deeper into above features of XGBoost.

Regularization

XGBoost as a member in Boosting family, is similar to GBM in many ways. First XGBoost fits an additive model of multiple base learner as GBM:

L2 Regularization

However, XGBoost add regularization into the model formulation(the loss function), which directly impact the training of each base learner.

$\hat{y} = \sum_k f_k $ is the current prediction

$\Omega(f) = \lambda T + \frac{1}{2} \lambda ||w||^2 $ where T is the number of leafs for base learner and W is the leaf assignment.

Shrinkage

L2 regularization prevents over-fitting by shrinking the parameter[ More detail discussed in this post ]. A more straightforward shrinkage method is directly using shrinkage factor to scale the output of each base learner, in order to reduce the impact of single base learner. The usage of shrinkage is same as the learning rate in neural network.

Column sampling

Also XGBoost brings in another Technic widely used in random forest - column sampling. This is similar to dropout rate in neural network, which spreads out the weights across features and also functions as bagging.

Relevant parameter

my_model = XGBRegressor(

    lambda=1, ## default=1, L2 Regularization

    alpha=0, ## default=0, L1 Regularization

    eta=0.3,  ## default=0.3, shrinkage rate

    colsample_bytree=1, ## default=1, % of column sampled for each base learner

    colsample_bylevel=1, ## default=1, % of column sampled for each split

)

Boosting

Compared with GBM, XGBoost uses 2nd order Taylor expansion to approximate the loss function. To make this more comparable to Tree - Gradient Boosting Machine with sklearn source code, let's first follow the same order: loss function, linear base learner and tree base learner.

Loss approximation - Newton Raphson

To approximate the loss function at t iteration, we do second order Taylor expansion at current prediction \(\hat{y}\), see below:

where \(g(\hat{y_i})\) is the gradient at current prediction, which is the one fitted in each base learner of GBM. And \(h(\hat{y_i})\) is the second order derivative, known as Hessian matrix in high dimension.

*Hessian Matrix

Hessian matrix is a square matrix of second order partial derivative, where \(H_ij = \frac{\partial^2f}{\partial{x_i} \partial{x_j}}\)

\(H = \begin{bmatrix}
\frac{\partial^2f}{\partial{x_1} \partial{x_1}}
& \frac{\partial^2f}{\partial{x_1} \partial{x_2}}
& \frac{\partial^2f}{\partial{x_1} \partial{x_3}} \\[0.3em] \frac{\partial^2f}{\partial{x_2} \partial{x_1}}
& \frac{\partial^2f}{\partial{x_2} \partial{x_2}}
& \frac{\partial^2f}{\partial{x_2} \partial{x_3}} \\[0.3em]
\frac{\partial^2f}{\partial{x_3} \partial{x_1}}
& \frac{\partial^2f}{\partial{x_3} \partial{x_2}}
& \frac{\partial^2f}{\partial{x_3} \partial{x_3}} \\[0.3em]
\end{bmatrix}\)

Given above example, we can tell the Hessian is symetric. And when f is convex, Hessian is positive semi-definite.

Linear base learner

When a linear base learner is used to optimize the loss function at each iteration. We can further simplify the above function:

Therefore Newton Raphson leads to a weighted least square regression against \(-\frac{g(\hat{y_i})}{h(\hat{y_i})}\) at each iteration.

In comparison, gradient descent is solving a least square regression against \(-g(\hat{y_i})\) at each iteration.

For linear base learner, gradient descent still need line search for optimal step length. While Newton Raphson solve the direction and step length at the same time. When \(h(\hat{y_i})\) is bigger, meaning \(g(\hat{y_i})\) change faster, then use smaller steps \(|-\frac{g(\hat{y_i})}{h(\hat{y_i})}|\).

Tree base learner

Let's add in L2 regularization to get full representation of loss function in XGBoost, and solve it with tree structure.

Using tree as base learner, all sample end up in the same leaf has same prediction and all leafs are disjoint. Therefore we can further simplify the loss function into:

A close solution of each leaf is below:

And following loss function:

However it is impossible to iterate through all possible tree structure to minimize above loss function. Therefore greedy search used in GBM is also applied here. We grow the tree from the root and search for best split at each iteration.

The best split is selected by maximum the loss reduction given above loss score, similar as Information Gain and Gini Index.

Where \(G_L = \sum_{x_i \in left}g(\hat{y_i})\) and \(H_L = \sum_{x_i \in left}h(\hat{y_i})\) is the sum of gradient/hessian in child node.

Pros and Cons

So what's the advantage of XGBoost over GBM?

1. Newton Raphson has better accuracy compared with Gradient Descent in estimating the loss reduction direction. Theoretically it should converge faster.

2. Regularization term \(\lambda\) helps prevent over-fitting by shrinking the weight on each individual base learner.

3. Regularization term \(\gamma\) can be viewed as a threshold for early stop. If the loss reduction is smaller than \(\gamma\) then stop growing. It prevents base learner from being too complicated.

However if Newton Raphson has so many advantages, then why gradient descent is more wildly used in machine learning. Newton Raphson also has its own limitation:

1. Computing Hessian matrix can be very time consuming

2. Loss function must have second order derivative

Relevant parameter

my_model = XGBRegressor(

    booster='gbtree', ## gbtree or gblinear

    max_depth=6, ## default=6, max depth of base learner.0 indicates no limit

    gamma=0  ## default=0, minimum loss reduction(threshold for early stopping)

)

Acceleration

One of the most time consuming part of Boosting algorithm is split finding, which consist of 2 parts: sorting features and searching through all values of each feature. Let's see how XGBoost optimizes these process.

Histogram binning - Approximate Algo

The key concept behind Boosting algorithm is to use biased simple base learner to approximate an unbiased final prediction. In other words accuracy is not the top concern for each base learner. Then we really don't need to search through all value to find the optimal split.

Therefore an approximate approach can be used. Basically each feature is split into buckets, and at each split only aggregate stats for each buckets are searched.

Here comes the next question. How many bins shall we use and how to bin the feature?

Number of bins is a hyper-parameter in the algorithm, the number of bin needs to be relative small compared with the unique feature value to speed up the algo. While too small number will lead to poor performance.

As for how to bin the feature, quantile is usually used to make sure data are evenly distributed. XGBoost propose weighted Quantile Sketch for binning, which is a weighted quantile with \(h(\hat{y_i})\) being the weight. This is in line with our analysis with Newton Raphson - it solves weighted least square for linear base learner.

Also you can choose to bin all the feature at the very beginning - Global method, or bin all the feature at each split - Local method. In comparison Global method needs more bins and only one computation, while Local method needs less bins but more computation, which may be better fit for deep tree.

Column Block

With histogram approximation, we no longer need to iterate across all values. What about feature sorting?

XGBoost use Column Block to solve this problem, where data is presorted for each column(feature) and stored in multiple blocks, which also allow parallel computation for split finding.

Parallel Computing

XGBoost is well known for its ability for parallel computing. Then which part of computation is parallelized?

1. Base learner is still trained one after another in series
2. Split finding for all feature runs in parallel
3. Split finding within one feature also runs in parallel

Sparsity aware

XGBoost provides a unified way to handle missing value, no matter it is manually created missing value (one hot encoding) or natural one. The mechanism has two benefit:

• Speed up sparse data training
  
  When search for optimal split for each feature, only consider non-missing value(0 or NA). \(I_k = \{i \in I| x_{ik} \notin \{0,NA\} \}\).
• improve missing value prediction
  
  Most of the algorithm need data cleaning to deal with missing value before training. We either remove or replace the missing value with some aggregate stats. For tree building, there is more options, missing value can following majority path, which has more observation.
  
  In XGBoost missing value is treated as a separate category. We calculate 2 optimal split by allowing missing value into left node and right node. And choose the direction that has bigger loss reduction.

Others

XGBoost has a few other designs to further speed up the algorithm.

Cache-aware Access prefetch the stats of gradient into buffer of each thread for exact algo. And optimize the block size to make sure the stats can fit into CPU cache for approximate algo.

Out-of Core Computation speed up disk reading by Block sharding (split data onto multiple disk) and Block compression (compressed data feature-wise and decompose on the fly) so that computation and disk reading can happen concurrently.

Relevant parameter

my_model = XGBRegressor(

    tree_method='auto', ## exact, approx, hist(optimized on approx with bin caching), gpu_exact, gpu_hist

    sketch_eps=6, ## default=0.03, lower eps leads to more bins(1 / sketch_eps). tree_method='approx'

    max_bin=0  ## default=256, tree_method='hist'

)

Reference

1. https://xgboost.readthedocs.io/en/latest/index.html
2. Chen, T. and Guestrin, C. (2016). Xgboost: A scalable tree boosting system. In Proceedings of the 22Nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16, pages 785–794, New York, NY, USA. ACM.
3. T. Hastie, R. Tibshirani and J. Friedman. Elements of Statistical Learning Ed. 2, Springer, 2009.