Answer to the problem no.

7
a) Error rate data without partitioning
50 50 50
𝐸𝑜𝑟𝑖𝑔 = 1 − max ( , )=
100 100 100

1. Splitting on attribute A, gain error rate

A=T A=F
+ 25 25
- 0 50
25 0 0
𝐸𝐴=𝑇 = 1 − 𝑚𝑎𝑥 ( , ) = =0
25 25 25
25 50 25
𝐸𝐴=𝐹 = 1 − 𝑚𝑎𝑥 ( , ) =
75 75 75
25 75 25
∆𝐴 = 𝐸𝑜𝑟𝑖𝑔 − 𝐸𝐴=𝑇 − 𝐸𝐴=𝐹 =
100 100 100

2. Splitting on attribute B, gain error rate

B=T B=F
+ 30 20
- 20 30
20
𝐸𝐵=𝑇 =
50
20
𝐸𝐵=𝐹 =
50
50 50 10
∆𝐵 = 𝐸𝑜𝑟𝑖𝑔 − 𝐸𝐵=𝑇 − 𝐸𝐵=𝐹 =
100 100 100

3. Splitting on attribute C, gain error rate

C=T C=F
+ 25 25
- 25 25
25
𝐸𝐶=𝑇 =
50
25
𝐸𝐶=𝐹 =
50
50 50 0
∆𝐶 = 𝐸𝑜𝑟𝑖𝑔 − 𝐸𝐶=𝑇 − 𝐸𝐶=𝐹 = =0
100 100 100
So choose the attribute A because it has the highest gain

b) The A=T child node is pure so no need for splitting, in A=F child node, the distribution
of training instances:
 Class label
B C
+ -
T T 0 20
F T 0 5
T F 25 0
F F 0 25

Classification error of node A=F:

25
𝐸𝑜𝑟𝑖𝑔 =
75
1. Splitting on attribute B, gain error rate
B=T B=F
+ 25 0
- 20 30
20
𝐸𝐵=𝑇 =
45
𝐸𝐵=𝐹 = 0
45 20 5
∆𝐵 = 𝐸𝑜𝑟𝑖𝑔 − 𝐸𝐵=𝑇 − 𝐸𝐵=𝐹 =
75 75 75

2. Splitting on attribute C, gain error rate

C=T C=F
+ 0 25
- 25 25
0
𝐸𝐶=𝑇 = =0
45
25
𝐸𝐶=𝐹 =
50
25 50
∆𝐶 = 𝐸𝑜𝑟𝑖𝑔 − 𝐸𝐶=𝑇 − 𝐸𝐶=𝐹 = 0
75 75

The split will be made on attribute B

c) 20 instances are misclassified (the error rate is 20/100)

d) C=T child node, the error rate before splitting is

25
𝐸𝑜𝑟𝑖𝑔 =
50

1. Splitting on attribute A, gain error rate

A=T A=F
+ 25 0
- 0 25
 𝐸𝐴=𝑇 = 0
𝐸𝐴=𝐹 = 0
25
∆𝐴 =
50

2. Splitting on attribute B, gain error rate

B=T B=F
+ 5 20
- 20 5
5
𝐸𝐵=𝑇 =
25
5
𝐸𝐵=𝐹 =
25
15
∆𝐵 =
50

Therefore, A is chosen as splitting attribute. For C=F child, the rate of error splitting is
25
𝐸𝑜𝑟𝑖𝑔 =
50
1. Splitting on attribute A, error rate is
A=T A=F
+ 0 25
- 0 25
𝐸𝐴=𝑇 = 0
25
𝐸𝐴=𝐹 =
50
∆𝐴 = 0

2. Splitting on attribute B, error rate is

B=T B=F
+ 0 25
- 0 25

𝐸𝐵=𝑇 = 0
𝐸𝐵=𝐹 = 0
25
∆𝐵 =
50
B is used as the splitting attribute. The overall error rate of the induced tree is 0.

e) The greedy heuristic does not necessarily lead to the best tree.