Today we will discuss more details about Regression Trees. This is the RT part of CART.

Before discussing Regression Trees, we will discuss a little more about Linear Regression.

The author uses the word forecasting and prediction interchangably.

As a Statistician, I have always tried to be specific about which word to use. Prediction is what is done when using Regression, or other similar methods, to predict a mean value or future observation within the range of the data. Forecasting is what is done with time series data when future observations are forecasted beyond the range of the data.

See wikipedia for further discussion.

• Prediction
• Forecasting

I would have titled the chapter Predicting Numeric Values - Regression Methods.

This is not that important in the big picture.

The simple linear regession model:

\(y_i = \beta_0 + \beta_1 x_i + \epsilon_i\)

where

\(\epsilon_i \sim N(0,\sigma_{\epsilon}^2)\)

Becuase of the distributional assumption on the error terms we can do Statistics. So we can test the statistical significance of the slope \(\beta_1\) and compute a confidence interval for the slope \(\beta_1\).

Parameter estimates are usually represented by the parameter with a hat. So the estimate of the slope in the simple linear regression would be

\(\hat{\beta}_1\)

The author introduces the a and b as the estimates.

The estimates are produced by minimizing the Sum of Squares Error for \(\beta_1\) and \(\beta_0\).

\(SSE = \sum (y_i - \hat{y}_i)^2 = \sum \hat{\epsilon}_i^2\)

where

\(\hat{y}_i = \hat{\beta_0} + \hat{\beta}_1 x_i\)

The estimates are

\(\hat{\beta}_1 = \frac{\sum(x_i - \bar{x})(y_i - \bar{y})}{\sum(x_i - \bar{x})^2} = r \cdot \frac{s_y}{s_x}\)

where

\(s_y^2 = \frac{\sum(y_i - \bar{y})^2}{n-1}\)

and

\(s_y = \sqrt{s_y^2}\)

and

\(\hat{\beta}_0 = \bar{y} - \hat{\beta}_1\cdot \bar{x}\)

So the fitted model would be

\(\hat{y}_i = \hat{\beta}_0 + \hat{\beta}_1 x_i\)

We could use this equation to draw a line on a scatterplot.

Question: What is the relationship between the correlation coefficient and the estimated slope coefficient?

Question: Can you define clearly what the correlation coefficient \(r\) measures?

Multiple Linear Regression includes more predictor variables.

Matrix notation is usually used to write down the model.

\(Y = X \cdot \beta + \epsilon\)

where

• \(Y\) is an \(n \times 1\) vector of the \(y_i\) values
• \(X\) is the \(n \times (p+1)\) design matrix
• \(\beta\) is an \((p+1) \times 1\) vector of the \(\beta_i\) values
• \(\epsilon\) is an \(n \times 1\) vector of the \(\epsilon_i\) values

The estimates are

\(\hat{\beta} = (X^T \cdot X)^{-1} X^T Y\)

• \(^{-1}\) is the inverse
• \(^T\) is the transpose

Trees for Numeric Prediction.

Strengths:

• trees for numeric data
• automatic feature selection
• no model in advance
• may work better than traditional regression
• does not require knowledge of statistics to intepret the results

Weaknesses:

• not so commonly used
• requires large amount of training data
• difficult to interpret effect of the predictors/features
• may not be as easy to interpret as a traditional regression model

Partitioning is done using a divide-and-conquer strategy according to the feature that will result in the greatest increase in homogeneity in the outcome after a split is preformed.

So the measurement is on the response variable/target variable \(Y\).

Standard deviation reduction (SDR)

\(SDR = sd(T) -\sum\frac{|T_i|}{|T|} \times sd(T_i)\)

where

\(sd(T)\) is the standard deviation of the \(Y\) values that are in the set \(T\).

\(|T|\) stands for the number of observations in the set \(T\).

Today we will look at the wine data example that tries to create a system to minic expert ratings of wine.

Wine Spectator's 100-point Scale

White Wine

Red Wine

Since we are not performing a Classification with Regression Trees, in this example, we cannot use a Confussion Matrix.

We will look at the correlation between the test values and the predicted values.

Another way to measure the error is to use the Mean Absolute Error (MAE)

\(MAE = \frac{1}{n} \sum |\epsilon_i|\)

or Mean Squared Error (MSE)

\(MSE = \frac{1}{n} \sum \epsilon^2_i\)