0% found this document useful (0 votes)
7 views2 pages

DATA7703 Week 7 Tutorial Solutions

The document provides solutions to a tutorial on classifiers, variance, and out-of-bag predictions in machine learning. It discusses the behavior of classifiers with different weights, variance calculations for combinations of random variables, and the implications of using bootstrap samples in model training. Additionally, it addresses the training error rates in relation to the number of basis models required for achieving zero training error.
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
7 views2 pages

DATA7703 Week 7 Tutorial Solutions

The document provides solutions to a tutorial on classifiers, variance, and out-of-bag predictions in machine learning. It discusses the behavior of classifiers with different weights, variance calculations for combinations of random variables, and the implications of using bootstrap samples in model training. Additionally, it addresses the training error rates in relation to the number of basis models required for achieving zero training error.
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

DATA7703 Week 7 Tutorial Solutions

2025 Semester 2
1. (a) No. Assume that 𝑤1 > 𝑤2 . From the table below, it is easy to see that for all the
four combinations of 𝑦1 , 𝑦2 ∈ {−1, +1}, the label 𝑦 = sgn(𝑤1 𝑦1 + 𝑤2 𝑦2 ) is always the
same as 𝑦1 .
𝑦1 𝑦2 𝑤1 𝑦1 + 𝑤2 𝑦2 𝑦
-1 -1 −𝑤1 − 𝑤2 -1
-1 1 −𝑤1 + 𝑤2 -1
1 -1 𝑤1 − 𝑤2 1
1 1 𝑤1 + 𝑤2 1
Thus we have 𝑐(𝑥) = sgn(𝑤1 𝑐1 (𝑥) + 𝑤2 𝑐2 (𝑥)) = 𝑐1 (𝑥) for any 𝑥. That is, 𝑐 is the same
as 𝑐1 when 𝑤1 > 𝑤2 . Similarly, 𝑐 is the same as 𝑐2 when 𝑤1 < 𝑤1 . So combining two
classifiers with different positive weights does not lead to a possibly more accurate
classifier.
If 𝑤1 = 𝑤2 , then sgn(𝑤1 𝑦1 + 𝑤2 𝑦2 ) = 0 when 𝑦1 + 𝑦2 = 0. That is, the combined pre-
diction is “uncertain”. While the combined classifier is different from both classifiers
in this case, it is again showing that combining two classifiers with positive weights
does not lead to a possibly more accurate classifier.
(b) The answer depends on whether 𝑐1 and 𝑐2 agree on any example 𝑥. If yes, simply
choose 𝑤1 = −1 and 𝑤2 = −2, then the prediction of 𝑐 on 𝑥 is opposite to those of
𝑐1 and 𝑐2 . If no, then 𝑐1 = −𝑐2 , and using the table in (a), it is easy to see that 𝑐 is
the same as either 𝑐1 or 𝑐2 .
2. (a) For any 𝑎, 𝑏, and any two random variables 𝑈, 𝑉 , we have Var(𝑎𝑈 +𝑏𝑉 ) = 𝑎2 Var(𝑈 )+
𝑏2 Var(𝑉 ) + 2𝑎𝑏 cov(𝑈, 𝑉 ) (verify this using definitions of variance and covariance if
you don’t know this). Hence
(︂ )︂
1 1 1 1 1
Var 𝑋1 + 𝑋2 = Var(𝑋1 ) + Var(𝑋2 ) + cov(𝑋1 , 𝑋2 )
2 2 4 4 2
1 1
= 𝜎 2 + 𝜌𝜎 2
2 2
1+𝜌 2
= 𝜎 .
2
(b) For any 𝑤1 , . . . , 𝑤𝑚 we have
𝑚
∑︁ ∑︁
Var(𝑤1 𝑋1 + . . . + 𝑤𝑚 𝑋𝑚 ) = 𝑤𝑖2 Var(𝑋𝑖 ) + 𝑤𝑖 𝑤𝑗 cov(𝑋𝑖 , 𝑋𝑗 )
𝑖=1 1≤𝑖̸=𝑗≤𝑚
𝑚
∑︁ ∑︁
= 𝑤𝑖2 𝜎 2 + 𝑤𝑖 𝑤𝑗 𝜌𝜎 2
𝑖=1 1≤𝑖̸=𝑗≤𝑚

1
1
When 𝑤1 , . . . , 𝑤𝑚 are all equal to 𝑚
, the last expression above is equal to
𝑚
∑︁ 1 2 ∑︁ 1 1 1
2
𝜎 + 2
𝜌𝜎 2 = 𝜎 2 + 𝑚(𝑚 − 1) 2 𝜌𝜎 2
𝑖=1
𝑚 1≤𝑖̸=𝑗≤𝑚
𝑚 𝑚 𝑚
1−𝜌 2
= 𝜌𝜎 2 + 𝜎 .
𝑚

3. (𝑥1 , 0) is an OOB observation for Tree 2 only, thus its OOB prediction is the one given
by Tree 2, that is, 1. Similarly, the OOB predictions for (𝑥2 , 1), (𝑥3 , 1), (𝑥4 , 0) are 1, 1,
1 respectively. Two out of the 4 OOB predictions are incorrect. Thus the OOB error is
2/4 = 0.5.

4. (a) No. All the models are built using bootstrap samples from the same dataset, and
thus they are correlated.
(b) No. Each time we construct a new model, the model is trained on a weighted training
set with weight determined by the predictions of the current ensemble. Thus a new
model depends on previous models, and is not independent of existing models.
(c) RF and bagging with DTs differs only in how a decision is trained on a bootstrap
sample. In bagging, the splitting variable at each node in a decision tree is chosen
from all 𝑑 features. In RF, the splitting variable at each node is chosen from 𝑚 < 𝑑
randomly chosen features — note that the random subset is separately sampled for
each node.

5. When 𝛾𝑡 ≥ 𝛾 > 0 for all 𝑡, the error rate of 𝐹𝑇 on the training set is upper bounded by
𝑇 √︀ 𝑇
2 2
∏︁ ∏︁ √︀
2
1 − 4𝛾𝑡 ≤ 1 − 4𝛾 2 = (1 − 4𝛾 2 )𝑇 /2 < 𝑒−4𝛾 ·𝑇 /2 = 𝑒−2𝛾 𝑇 ,
𝑡=1 𝑡=1

where we make use of the inequality 1 − 𝑥 < 𝑒−𝑥 for 𝑥 > 0.


Since there are 𝑛 training examples, an error rate less than 𝑛1 is the same as zero training
2
error. Thus if 𝑒−2𝛾 𝑇 ≤ 𝑛1 , then the training error is zero. Solving the inequality for 𝑇 ,
we have 𝑇 ≥ ln 𝑛
2𝛾 2
. That is, as long as 𝑇 ≥ ln 𝑛
2𝛾 2
, we have zero training error. Equivalently,
ln 𝑛
at most 2𝛾 2 basis models are needed to make the training error zero (more precisely, at
most ln 𝑛
2𝛾 2
− 1 basis models are needed).

You might also like