*67*

In the field of machine learning, our goal is to build a model that can effectively use a set of predictor variables to predict the value of some response variable.

Given a set of *p* total predictor variables, there are many models that we could potentially build. One method that we can use to pick the best model is known as best subset selection, which attempts to choose the best model from *all* possible models that could be built with the set of predictors.

Unfortunately this method suffers from two drawbacks:

- It can be computationally intense. For a set of
*p*predictor variables, there are 2^{p}possible models. For example, with 10 predictor variables there are 2^{10}= 1,000 possible models to consider. - Because it considers such a large number of models, it could potentially find a model that performs well on training data but not on future data. This could result in overfitting.

An alternative to best subset selection is known as **stepwise selection**, which compares a much more restricted set of models.

There are two types of stepwise selection methods: forward stepwise selection and backward stepwise selection.

**Forward Stepwise Selection**

Forward stepwise selection works as follows:

**1.** Let M_{0} denote the null model, which contains no predictor variables.

**2.** For k = 0, 2, … p-1:

- Fit all p-k models that augment the predictors in M
_{k}with one additional predictor variable. - Pick the best among these p-k models and call it M
_{k+1}. Define “best” as the model with the highest R^{2}or equivalently the lowest RSS.

**3.** Select a single best model from among M_{0}…M_{p} using cross-validation prediction error, Cp, BIC, AIC, or adjusted R^{2}.

**Backward Stepwise Selection**

Backward stepwise selection works as follows:

**1.** Let M_{p} denote the full model, which contains all *p* predictor variables.

**2.** For k = p, p-1, … 1:

- Fit all k models that contain all but one of the predictors in M
_{k}, for a total of k-1 predictor variables. - Pick the best among these k models and call it M
_{k-1}. Define “best” as the model with the highest R^{2}or equivalently the lowest RSS.

**3.** Select a single best model from among M_{0}…M_{p} using cross-validation prediction error, Cp, BIC, AIC, or adjusted R^{2}.

**Criteria for Choosing the “Best” Model**

The last step of both forward and backward stepwise selection involves choosing the model with the lowest prediction error, lowest Cp, lowest BIC, lowest AIC, or highest adjusted R^{2}.

Here are the formulas used to calculate each of these metrics:

**Cp:** (RSS+2dσ̂) / n

**AIC: **(RSS+2dσ̂^{2}) / (nσ̂^{2})

**BIC: **(RSS+log(n)dσ̂^{2}) / n

**Adjusted R ^{2}:** 1 – ( (RSS/(n-d-1)) / (TSS / (n-1)) )

where:

**d:**The number of predictors**n:**Total observations**σ̂:**Estimate of the variance of the error associate with each response measurement in a regression model**RSS:**Residual sum of squares of the regression model**TSS:**Total sum of squares of the regression model

**Pros & Cons of Stepwise Selection**

Stepwise selection offers the following **benefit**:

It is more computationally efficient than best subset selection. Given *p* predictor variables, best subset selection must fit 2^{p} models.

Conversely, stepwise selection only has to fit 1+p(p+ 1)/2 models. For p = 10 predictor variables, best subset selection must fit 1,000 models while stepwise selection only has to fit 56 models.

However, stepwise selection has the following potential **drawback:**

It is not guaranteed to find the best possible model out of all 2^{p} potential models.

For example, suppose we have a dataset with p = 3 predictors. The best possible one-predictor model may contain x_{1} and the best possible two-predictor model may instead contain x_{1} and x_{2}.

In this case, forward stepwise selection will fail to select the best possible two-predictor model because M_{1} will contain x_{1}, so M_{2} must also contain x_{1} along with some other variable.