decision trees ii and bias/variance and cross-validation

Decision Trees II and Bias/Variance andCross-Validation

Nipun Batra and teaching staffJanuary 11, 2020

IIT Gandhinagar

Real Input Real Output

Example 1

Let us consider the dataset given below

1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0Feature, X

0

1

2

Labe

l, Y

1

Example 1

What would be the prediction for decision tree with depth 0?

1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0Feature, X

0

1

2

Labe

l, Y

2

Example 1

Prediction for decision tree with depth 0.Horizontal dashed line shows the predicted Y value. It is theaverage of Y values of all datapoints.

1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0Feature, X

0

1

2

Labe

l, Y

3

Example 1

What would be the decision tree with depth 1?

1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0Feature, X

0

1

2

Labe

l, Y

4

Example 1

Decision tree with depth 1

X[0] <= 2.5mse = 0.667samples = 6value = 1.0

mse = 0.0samples = 2value = 0.0

True


False

5

Example 1

The Decision Boundary

1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0Feature, X

0

1

2La

bel,

Y

6

Example 1

What would be the decision tree with depth 2 ?

1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0Feature, X

0

1

2

Labe

l, Y

7

Example 1

Decision tree with depth 2



True


False



8

Example 1

The Decision Boundary

1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0Feature, X

0

1

2La

bel,

Y

9

Objective function

Here, Feature is denoted by X and Label by Y.Let the “decision boundary” or “split” be at X = S.Let the region X < S, be region R1.Let the region X > S, be region R2.

10

Objective function


Then, letC1 = Mean (Yi|Xi ∈ R1)C2 = Mean (Yi|Xi ∈ R2)

10

Objective function


Then, letC1 = Mean (Yi|Xi ∈ R1)C2 = Mean (Yi|Xi ∈ R2)Loss =

∑i((Yi − C1|Xi ∈ R1)2 + (Yi − C2|Xi ∈ R2)2)

10

How to find optimal split “S”?

11


1. Sort all datapoints (X,Y) in increasing order of X.

12


1. Sort all datapoints (X,Y) in increasing order of X.

2. Evaluate the loss function for all

S =Xi+Xi+1

2

and the select the S with minimum loss.

12

A Question!

Draw a regression tree for Y = sin(X), 0 ≤ X ≤ 2π

13

A Question!

Dataset of Y = sin(X), 0 ≤ X ≤ 7 with 10,000 points

0 1 2 3 4 5 6 7Feature, X

1

0

1La

bel,

Y

14

A Question!

Regression tree of depth 1

X[0] <= 3.033mse = 0.463

samples = 10000value = 0.035

mse = 0.086samples = 4333

value = 0.657

True

mse = 0.23samples = 5667value = -0.441

False

15

A Question!

Decision Boundary

1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0Feature, X

0

1La

bel,

Y

16

A Question!

Regression tree with no depth limit is too big to fit in a slide.It has of depth 20. The decision boundaries are in figure below.

1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0Feature, X

0

1La

bel,

Y

17

Summary

• Interpretability an important goal• Decision trees: well known interpretable models• Learning optimal tree is hard• Greedy approach:• Recursively split to maximize “performance gain”• Issues:

• Can overfit easily!• Empirically not as powerful as other methods

18

A Question!

What would be the decision boundary of a decision treeclassifier?

1 2 3 4 5 6 7X1

1.0

1.5

2.0

2.5

3.0

3.5

4.0

X 2

19

Decision Boundary for a tree with depth 1

1 2 3 4 5 6 7X1

1.0

1.5

2.0

2.5

3.0

3.5

4.0

X 2

(a) Decision Boundary

X[0] <= 4.0entropy = 1.0samples = 12value = [6, 6]

entropy = 0.65samples = 6value = [5, 1]

True


False

(b) Decision Tree

20

Decision Boundary for a tree with no depth limit

1 2 3 4 5 6 7X1

1.0

1.5

2.0

2.5

3.0

3.5

4.0

X 2

(c) Decision Boundary



True


False


X[0] <= 1.5entropy = 0.918

samples = 3value = [2, 1]




X[0] <= 6.5entropy = 0.918




(d) Decision Tree

21

Are deeper trees always better?

As we saw, deeper trees learn more complex decisionboundaries.

22

Are deeper trees always better?

As we saw, deeper trees learn more complex decisionboundaries.

But, sometimes this can lead to poor generalization

22

An example

Consider the dataset below

1 2 3 4 5 6X1

1

2

3

4

5

6

X 2

Train dataset

(e) Train Set

1 2 3 4 5 6X1

1

2

3

4

5

6

X 2

Test dataset

(f) Test Set

23

Underfitting

Underfitting is also known as high bias, since it has a verybiased incorrect assumption.

1 2 3 4 5 6X1

1

2

3

4

5

6

X 2

(g) Decision Boundary

X[1] <= 1.5entropy = 0.964samples = 36

value = [22, 14]


True


value = [16, 14]

False

entropy = 0.98samples = 24

value = [10, 14]


(h) Decision Tree

24

Overfitting

Overfitting is also known as high variance, since very smallchanges in data can lead to very different models.Decision tree learned has depth of 10.

1 2 3 4 5 6X1

1

2

3

4

5

6

X 2

25

Intution for Variance

A small change in data can lead to very different models.

Dataset 1

1 2 3 4 5 6X1

1

2

3

4

5

6

X 2

Train dataset

Dataset 2

1 2 3 4 5 6X1

1

2

3

4

5

6

X 2

26

Intution for Variance


value = [22, 14]


True


value = [16, 14]

False


value = [10, 14]








X[1] <= 2.5entropy = 0.811




X[1] <= 4.5entropy = 0.918











value = [22, 14]


True


value = [16, 14]

False


value = [10, 14]






X[1] <= 2.5entropy = 0.811






X[0] <= 4.5entropy = 0.918




27

A Good Fit

1 2 3 4 5 6X1

1

2

3

4

5

6

X 2


value = [22, 14]


True


value = [16, 14]

False


value = [10, 14]






28

Accuracy vs Depth Curve

0 2 4 6 8 10Depth

60

70

80

90

100Ac

cura

cy

Test AccuracyTrain Accuracy

As depth increases, train accuracy improvesAs depth increases, test accuracy improves till a pointAt very high depths, test accuracy is not good (overfitting).

29


0 2 4 6 8 10Depth

60

70

80

90

100Ac

cura

cy


As depth increases, train accuracy improves

As depth increases, test accuracy improves till a pointAt very high depths, test accuracy is not good (overfitting).

29


0 2 4 6 8 10Depth

60

70

80

90

100Ac

cura

cy


As depth increases, train accuracy improvesAs depth increases, test accuracy improves till a point

At very high depths, test accuracy is not good (overfitting).

29


0 2 4 6 8 10Depth

60

70

80

90

100Ac

cura

cy


As depth increases, train accuracy improvesAs depth increases, test accuracy improves till a pointAt very high depths, test accuracy is not good (overfitting). 29

Accuracy vs Depth Curve : Underfitting

The highlighted region is the underfitting region.Model is too simple (less depth) to learn from the data.

2 4 6 8 10Depth

60

70

80

90

100

Accu

racy


30

Accuracy vs Depth Curve : Overfitting

The highlighted region is the overfitting region.Model is complex (high depth) and hence also learns theanomalies in data.

2 4 6 8 10Depth

60

70

80

90

100

Accu

racy


31


The highlighted region is the good fit region.We want to maximize test accuracy while being in this region.

2 4 6 8 10Depth

60

70

80

90

100

Accu

racy


32

The big question!?

How to find the optimal depth for a decision tree?

33

The big question!?

How to find the optimal depth for a decision tree?

Use cross-validation!

33

Our General Training Flow

Train Data Test Data

Model

Training

Predicting

Actual Labels

Predicted Labels

Error MetricCalculation

Accuracy

34

K-Fold cross-validation: Utilise full dataset for testing

Train Test

Train Test Train

Train Test Train

Test Train

FOLD 1

FOLD 2

FOLD 3

FOLD 4

35

The Validation Set

Train Data Test Data

ModelDepth = 1

Training

Predicting

Actual Labels

Predicted Labels

Error MetricCalculation

Accuracy

Validation

ModelDepth = 2

ModelDepth = 3

Accuracy on validation set

Validation AccuracyDepth 1 : 70%Depth 2 : 90%Depth 3 : 80%

Select model having highest validation

accuracy

36

Nested Cross Validation

Divide your training set into K equal parts.Cyclically use 1 part as “validation set” and the rest for training.Here K = 4

Train Validation

Train Validation Train


Validation Train

FOLD 1

FOLD 2

FOLD 3

FOLD 4

37

Nested Cross Validation

Average out the validation accuracy across all the foldsUse the model with highest validation accuracy

Train Validation



Validation Train

FOLD 1

FOLD 2

FOLD 3

FOLD 4

Average out Validation Accuracy across all folds.

Select the hyperparameter for which the model gives the highest average validation accuracy

38

Next time: Ensemble Learning

• How to combine various models?• Why to combine multiple models?• How can we reduce bias?• How can we reduce variance?

39

decision trees ii and bias/variance and cross-validation

Documents