
# Gradient Boosting In Forests

Our journey has taken us to the most powerful out-of-the-box machine learning algorithm: gradient-boosted ensemble models. We've already seen one ensemble model: Random forests, which is an ensemble of decision trees. However the next algorithm under review today outperforms random forests on all machine-learning league tables: **gradient boosted trees**. This is always worth pulling out of your back pocket and applied first to new data problems. Why? It's performance and ease of use is simply unparalleled. In this tutorial we'll begin by gently unpacking the theory behind the algorithm.

<Callout type="success" heading={"📑 Learning Objectives"}>
- Introduce the notion of _Gradient Boosted Trees_.
- Explain terminology: 'weak learners', 'stumps', 'residuals', 'gradients'.
- Write down the mathematical trick that drives the algorithm.
- Walk through the algorithm in its most basic form.

</Callout>

<Image
	src={
		"/images/courses/machine-learning-models-for-professionals/give-yourself-a-gradient-boost/img1.jpg"
	}
	alt={"Shining a Light on the Stumps"}
	width={910}
	height={607}
/>

## From trees to stumps with autocorrection

Our journey has brought us from trees to forests and out the other side, a field full of stumps. When considering a collection of random forests, we decided that they should be trained independently, and thus also _simultaneously_. Then the final answer would be formed by averaging over all the outputs in the forest. This method required us to grow trees in proportion to the size (resp. square-root of the size) of the data set for regression problems (resp. classification).

<Callout type="info" heading={""}>
	Gradient boosted trees are trained in sequence, each dependent on the previous
	tree: **each tree learns the _error_ in the previous prediction** (the
	_residual_ in regression problems), rather than recreating the data itself.
	This means, if a data point prediction is miss-predicted by the model, the next
	tree shall divert attention there. At the same time, all correct predictions
	would have a small error and be grouped together around zero.
</Callout>

This is so effective, we are able to reduce the size of the tree to a **stump**: a decision tree with a single splitting. These small intermediate models, learning residuals, are formally known as 'weak learners'. Furthermore, we aren't tied to using just tree based-models. The core concept works well with any model.

## Walking though the algorithm with a visual example

Let's define the algorithm by example: a simple quadratic curve, $y(x) = x^2.$ For comparison, let's first reconstruct a random forest of stumps to see how it fares.

Our training data shall consist of 10 points on this curve. Let's set up an example. Our goal is to interpolate between these points with 100 test points.

The training target is then the following list of numbers:

$$
\mathbf{y} = [1.00, 0.60, 0.31, 0.11, 0.01, 0.01, 0.11, 0.31, 0.60, 1.00]
$$

### A random forest of stumps

First, let's construct our base comparative model: a random forest of stumps. We'll train each stump in the forest on bagged data (recall this definition)--on the contrary, the gradient boosted tree shall be trained on the entire data set.

This leaves us with a list of models, the 'stumps' which all make predictions via a single splitting. To form the aggregate model (the random forest), as this is a regression problem we take the mean of the predictions.

<Image
	src={
		"/images/courses/machine-learning-models-for-professionals/auto-correction-in-a-forest-of-stumps/img1.png"
	}
	alt={"A Random Forest of Stumps"}
	width={568}
	height={453}
	caption={
		"Figure 1. Using a collection of shallow trees, 'stumps', to form a random forest does not have enough flex to solve simple problems."
	}
/>

This is not an impressive model. The locations of the splittings impact the model drastically. The shape of the error function forces those splitting toward the sides leaving a void in the centre. Averaging a collection of stumps is simply too coarse for this data set.

Let's check the Mean-Squared Error (MSE) for each stump, as well as the forest.

|          | Stump 1 | Stump 2 | Stump 3 | Stump 4 | Mean  |
| :------: | :-----: | :-----: | :-----: | :-----: | :---: |
| **MSE:** |  0.074  |  0.194  |  0.194  |  0.201  | 0.042 |

The error of each stump is almost five times as high, and uniform across the stumps. We shall keep this metric in mind as we move to our next model.

### Out of the woods with gradient boosting

Now let's walk through the gradient boosting algorithm. Each stump is dependent on the previous, and hence each previous. We shall denote our dataset by $(X, \mathbf{y})$, where we are trying to predict the vector $\mathbf{y}$ from the columns of $X$.

1. Let's call the first stump $T_0$. This model shall be a straight forward decision tree, trained on the original set $(X, \mathbf{y})$.

2. Compute the residual error $R_1 = \mathbf{y} - T_0$

$$
R_1 = [0.00, 0.26, -0.03, -0.23, -0.32, -0.33, -0.23, -0.03,  0.26,  0.66]
$$

<Callout type="warning" heading={""}>
	💭 The MSE for $T_0$ is the sum of the squares of numbers in $R_1$. Keep this
	in mind for later.
</Callout>

3. Now we train the next tree stump $T_1$ on these residuals, so that the training data is $(X, R_1)$. We are now modeling the difference between $T_0$ and the target $\mathbf{y}$.

4. Note that $T_0 + T_1 \approx T_0 + \mathbf{y} - T_0 = \mathbf{y}$. This shall be our ensemble model. And the next residual is formed by the difference
   $
R_2 = \mathbf{y} - (T_0 + T_1) = R_1 - T_1
$
   so that

$$
R_2 = [0.12,  0.38,  0.08, -0.12, -0.21, -0.21, -0.12,  0.08, -0.20,  0.20]
$$

5. Now we can iterate this to a given number of stumps ($N=100$ say), where the $i$-th stump $T_i$ is is trained by using $X$ to predict the previous residuals $R_i = R_{i-1} - T_{i-1}$ at which stage the model is evaluated by $\sum_{i=0}^{n} T_{i}$.

In our example, $N=4$ and we can compare the final model to the output of the random forest.

<Image
	src={
		"/images/courses/machine-learning-models-for-professionals/auto-correction-in-a-forest-of-stumps/img2.png"
	}
	alt={"Gradient Boosting Stumps"}
	width={568}
	height={453}
	caption={
		"Figure 2. Applying Gradient Boosting to the stumps immediately shows far better convergence with the small number of parameters available."
	}
/>

By eye, it seems that the performance is very much better (see the fourth solid line). To be sure, let's numerically assess the performance.

|     | 1 Stump | 2 Stumps | 3 Stumps | 4 Stumps |
| :-: | ------: | -------: | -------: | -------: |
| MSE |   0.074 |    0.039 |    0.022 |    0.019 |

As expected, in contrast to the independent random forest stumps, the MSE is decreasing. And with 4 stumps, the reported MSE of 0.019 is half as much as the MSE for the mean prediction of the forest.

<Callout type="info" heading={""}>
**More general gradients**
This construction is theoretically delightful. If were to impose a mean squared loss on our model,

$$
L(T_n) = \frac{1}{2N}\sum_{i=1}^{N}\|T_n - R_n\|^2,
$$

then with a gradient boosted model $F_n=\sum_{i=1}^{n} T_i$, the gradient of this loss is precisely equal to the residuals:

$$
\frac{\partial L(T_{n})}{\partial T_n} = R_{n+1}
$$

and so we could pick a small real number $\alpha$, a _learning rate_ and perform our model update $F_n\mapsto F_{n+1} = F_{n} + \alpha T_{n+1}$, scaled by the learning rate, we recover the update formula used in gradient decent. And so this algorithm fits into a common optimisation framework. We can then also justify taking more general loss functions $L$ to use on bespoke problems.

</Callout>


In this lesson we take a first principles look at the algorithm underpinning Gradient Boosted Trees: how does one boost a gradient? We can take hints from this neat mathematical trick when building our own models, and importantly understand what makes XGBoost so successful.

Please log in or subscribe to continue