Math 185 Final Project (Due December 8)
Problem 1
The baseball dataset consists of the statistics of 263 players in Major League
Baseball in the season 1986. The dataset (hitters.csv) consist of 20 variables:
Variable Description
AtBat Number of times at bat in 1986
Hits Number of hits in 1986
HmRun Number of home runs in 1986
Runs Number of runs in 1986
RBI Number of runs batted in in 1986
Walks Number of walks in 1986
Years Number of years in major leagues
CAtBat Number of times at bat during his career
CHits Number of hits during his career
CHmRun Number of home runs during his career
CRuns Number of runs during his career
CRBI Number of runs batted in during his career
CWalks Number of walks during his career
League A factor with levels A (coded as 1) and N (coded as 2) indicating
player’s league at the end of 1986
Division A factor with levels E (coded as 1) and W (coded as 2) indicating
player’s division at the end of 1986
PutOuts Number of put outs in 1986
Assists Number of assists in 1986
Errors Number of errors in 1986
Salary 1987 annual salary on opening day in thousands of dollars
NewLeague A factor with levels A (coded as 1) and N (coded as 2) indicating
player’s league at the beginning of 1987
In this problem, we use Salary as the response variable, and the rest 19 variables
as predictors/covariates, which measure the performance of each player in season
1986 and his whole career. Write R functions to perform variable selection using
best subset selection partnered with BIC (Bayesian Information Criterion):
1) Starting from the null model, apply the forward stepwise selection algorithm to
produce a sequence of sub-models iteratively, and select a single best model
using the BIC. Plot the “BIC vs Number of Variables” curve. Present the selected
model with the corresponding BIC.
2) Starting from the full model (that is, the one obtained from minimizing the
MSE/RSS using all the predictors), apply the backward stepwise selection
algorithm to produce a sequence of sub-models iteratively, and select a single
best model using the BIC. Plot the “BIC vs Number of Variables” curve. Present
the selected model with the corresponding BIC.
3) Are the selected models from 1) and 2) the same?
Problem 2
In this problem, we fit ridge regression on the same dataset as in Problem 1. First,
standardize the variables so that they are on the same scale. Next, choose a grid of
𝜆 values ranging from 𝜆 = 1010 to 𝜆 = 10−2
, essentially covering the full range of
scenarios from the null model containing only the intercept, to the least squares fit.
For example:
> grid = 10^seq(10, -2, length=100)
1) Write an R function to do the following: associated with each value of 𝜆 ,
compute a vector of ridge regression coefficients (including the intercept),
stored in a 20 × 100 matrix, with 20 rows (one for each predictor, plus an
intercept) and 100 columns (one for each value of 𝜆).
2) To find the “best” 𝜆 , use ten-fold cross-validation to choose the tuning
parameter from the previous grid of values. Set a random seed – set.seed(1),
first so your results will be reproducible, since the choice of the cross-validation
folds is random. Plot the “Cross-Validation Error versus 𝜆” curve, and report the
selected 𝜆.
3) Finally, refit the ridge regression model on the full dataset, using the value of 𝜆
chosen by cross-validation, and report the coefficient estimates.
Remark: You should expect that none of the coefficients are zero – ridge regression
does not perform variable selection.
Problem 3
In this problem, we revisit the best subset selection problem. Given a response
vector 𝑌 = (𝑦1, … , 𝑦𝑛)𝑇
and an 𝑛 × 𝑝 design matrix 𝑋 = (𝑥1, … , 𝑥𝑛)𝑇 with 𝑥𝑖 =(𝑥𝑖1, … , 𝑥𝑖𝑝)𝑇. For 1 ≤ 𝑘 ≤ 𝑝, let 𝛽̂0, 𝛽̂ be the solution to the following sparsityconstrained
least squares problem:
Based on the property 𝛽̂0 = 𝑦̅ − 𝑥̅𝑇𝛽̂, we can center 𝑌 and 𝑋 first to get rid of the
intercept,
where 𝑌̃ and 𝑋̃ represent the centered 𝑌 and 𝑋, respectively. To solve this, we
introduce the Gradient Hard Thresholding Pursuit (GraHTP) algorithm. Let 𝑓(𝛽) =
∥ 𝑌̃ − 𝑋̃𝛽 ∥2
2⁄(2𝑛) be the objective function.
GraHTP Algorithm.
Input: 𝑌̃, 𝑋̃, sparsity 𝑘, stepsize 𝜂 > 0
(Hint: normalize the columns of 𝑋̃ to have variance 1).
Initialization: 𝛽
0 = 0, 𝑡 = 1.
repeat
1) Compute 𝛽̃𝑡 = 𝛽𝑡−1 − 𝜂∇𝑓(𝛽𝑡−1);
2) Let 𝒮𝑡 = supp(𝛽̃𝑡, 𝑘) be the indices of 𝛽̃𝑡 with the largest 𝑘 absolute values;
3) Compute 𝛽𝑡 = argmin{𝑓(𝛽); supp(𝛽) ⊆ 𝒮𝑡};𝑡 = 𝑡 + 1;
until convergence, i.e. ∥ 𝛽𝑡 − 𝛽𝑡−1∥2< 10−4.
Output: 𝛽𝑡.
1) Write an R function to implement the above GraHTP algorithm.
2) Consider again the baseball dataset in Problem 1 with 𝑛 = 263, 𝑝 = 19. For 𝑘 =
1, … , 𝑝, use the above function to find the best 𝑘-sparse model, denoted by ℳ𝑘
.Then use BIC to select a single best model among ℳ1 … ,ℳ𝑝.
3) Compare your result with those obtained in Problem 1.