{[ promptMessage ]}

Bookmark it

{[ promptMessage ]}

lecture_04

# lecture_04 - Financial Data Mining FSRM588 Lecture 04...

This preview shows pages 1–10. Sign up to view the full content.

Financial Data Mining FSRM588 Lecture 04: Penalized Least Squares & Principal Component Analysis Department of Statistics & Biostatistics Rutgers University October 01 2013

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Outline 1 Penalized Least Squares 2 Principal Component Analysis
Lasso Lasso solves the following optimization problem, ˆ β lasso = arg max β R p +1 - 1 2 N N X i =1 y i - β 0 - p X j =1 x ij β j 2 - λ p X j =1 | β j | for some λ 0 .

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Penalized Least Squares We can use some other penalty on the parameters β j as well, and consider the following general penalized least squares problem ˆ β = arg max β R p +1 - 1 2 N N X i =1 y i - β 0 - p X j =1 x ij β j 2 - p X j =1 p λ ( | β j | ) , where p λ ( · ) is the penalty function. Best subset selection: p λ ( t ) = ( λ 2 / 2) I ( t 6 = 0) . Lasso: p λ ( t ) = λ | t | . Hard thresholding: p λ ( t ) = 1 2 [ λ 2 - ( λ - t ) 2 + ] . Elastic net: p λ ( t ) = λ ± at 2 + (1 - a ) | t | ² with 0 a 1 . Bridge regression: p λ ( t ) = λ | t | q for some 0 < q 2 . Ridge regression: p λ ( t ) = λ | t | 2 . SCAD: to be introduced.
L 1 and SCAD 0 1 2 3 4 5 6 0 2 3 4 5 6 Penalty Functions t penalty L1 SCAD Hard 0 1 2 3 4 5 6 0.0 0.5 1.0 1.5 2.0 Derivatives t derivative L1 SCAD Hard

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Centering Usually the intercept is not penalized, so we center y and x j ﬁrst and consider the following optimization problem ˆ β = arg max β R p - 1 2 N N X i =1 y i - p X j =1 x ij β j 2 - p X j =1 p λ ( | β j | ) , with the implicit assumption that 1 0 x j = 0 , 1 j p and 1 0 y = 0 .
Canonical Model Consider the special case by assuming the columns of the input matrix N - 1 / 2 X are orthonormal, i.e. X 0 X /N = I . Let ˆ β ls j = N - 1 x 0 j y , , ˆ y ls = XX 0 y , the penalized least squares can be rewritten as ˆ β = arg max β R p - 1 2 N ± ± y - ˆ y ls ± ± 2 + p X j =1 ² - 1 2 ³ ˆ β ls j - β j ´ 2 - p λ ( | β j | ) µ , so that the optimization problem becomes maximizing for each β j , ˆ β j = arg max β j R - 1 2 ³ ˆ β ls j - β j ´ 2 + p λ ( | β j | ) · . (1)

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Desired Properties Lasso estimates are ˆ β lasso j = sign( ˆ β ls j )( | ˆ β ls j | - λ ) + . Consider a simpliﬁed case of. Assume we have N training points ( x 1 ,y 1 ) ,..., ( x N N ) from the following model y i = β 1 x i 1 + β 2 x i 2 + ± i , with β 1 > 0 and β 2 = 0 , and ± i are i.i.d. with mean zero and variance σ 2 . Desired Properties 1 Sparsity. As N goes to inﬁnity, β 2 is estimated as zero with probability approaching one. 2 Unbiasedness. β 1 can be estimated with small bias. More speciﬁcally, we would like the estimate ˆ β 1 to have the property N ( ˆ β 1 - β 1 ) ⇒ N (0 2 ) .
Lasso Fails Rewrite λ N = N - 1 / 2 η N . From the identity P ( ˆ β Lasso 2 = 0) = P ± | ˆ β ls 2 | ≤ N - 1 / 2 η N ² = P ± | N ˆ β ls 2 | ≤ η N ² , we see that in order for this probability to approaching one, we must let the sequence η N diverge as N grows. On the other hand, N ± ˆ β lasso 1 - β 1 ² N ± ˆ β ls 1 - λ N - β 1 ² = N ± ˆ β ls 1 - β 1 ² - η N , from which we ﬁnd that in order for (2) to hold, the sequence η N must converge to zero. This simple example tells us that Lasso

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
This is the end of the preview. Sign up to access the rest of the document.

{[ snackBarMessage ]}

### Page1 / 32

lecture_04 - Financial Data Mining FSRM588 Lecture 04...

This preview shows document pages 1 - 10. Sign up to view the full document.

View Full Document
Ask a homework question - tutors are online