Scribd
Upload
272 views

Random Forest

This document provides an introduction to random forests, which are an ensemble machine learning technique. It begins with an outline and then covers decision trees, bagging, and how random forests combine these methods by growing many randomized decision trees and aggregating their predictions. Key points include how random forests encourage diversity among trees to improve accuracy over a single tree, and how they produce useful byproducts like out-of-bag error estimates and variable importance measures. The document also briefly introduces kernel-induced random forests, which apply the random forest approach to data in a kernel-induced feature space.

Uploaded by

azn
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
272 views

Random Forest

This document provides an introduction to random forests, which are an ensemble machine learning technique. It begins with an outline and then covers decision trees, bagging, and how random forests combine these methods by growing many randomized decision trees and aggregating their predictions. Key points include how random forests encourage diversity among trees to improve accuracy over a single tree, and how they produce useful byproducts like out-of-bag error estimates and variable importance measures. The document also briefly introduces kernel-induced random forests, which apply the random forest approach to data in a kernel-induced feature space.

Uploaded by

azn
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 30

An introduction

to random forests
Eric Debreuve / Team Morpheme
Institutions: University Nice Sophia Antipolis / CNRS / Inria
Labs: I3S / Inria CRI SA-M / iBV

Outline
Machine learning
Decision tree
Random forest
Bagging
Random decision trees

Kernel-Induced Random Forest (KIRF)


Byproducts
Out-of-bag error
Variable importance
2

Machine learning
Learning/training: build a classification or regression rule
from a set of samples
Samples
(learning set)

Machine learning
algorithm

class/category = rule(sample)
or
value = rule(sample)

Prediction: assign a class or value to new samples


New samples

Learned rule

predicted class
or
predicted value

(Un)Supervised learning
Supervised
Learning set = { (sample

[acquisition],

class

[expert])

Unsupervised
Learning set = unlabeled samples

Semi-supervised
Learning set = some labeled samples + many unlabeled samples

Ensemble learning
Combining weak classifiers (of the same type)...
... in order to produce a strong classifier
Condition: diversity among the weak classifiers

Example: Boosting
Train each new weak classifier focusing on samples misclassified by
previous ones
Popular implementation: AdaBoost
Weak classifiers: only need to be better than random guess

Outline
Machine learning
Decision tree
Random forest
Bagging
Random decision trees

Kernel-Induced Random Forest (KIRF)


Byproducts
Out-of-bag error
Variable importance
6

Decision tree
Root node

Q1

Entry point to a collection of data

Inner nodes (among which the root node)


A question is asked about data
One child node per possible answer

Q2
D1

D2

Q3
D3

D4

D5

Leaf nodes
Correspond to the decision to take (or conclusion to make) if reached

Example: CART - Classification and Regression Tree


Labeled sample
Vector of variable/feature values + class label
Binary decision tree
Top-down, greedy building...
... by recursively partitioning the feature space into hyper-rectangles
Similarity with weighted kNN

Normally, pruning
To avoid over-fitting of learning data
To achieve a trade-off between prediction accuracy and complexity
7

Decision tree > CART > Building


All labeled samples initially assigned to root node
N root node
With node N do
Find the feature F + threshold value T...
... that split the samples assigned to N into 2 subsets Sleft and Sright...
... so as to maximize the label purity within these subsets
Assign (F,T) to N
If Sleft and Sright too small to be splitted
Attach child leaf nodes Lleft and Lright to N
Tag the leaves with the most present label in Sleft and Sright, resp.
else
Attach child nodes Nleft and Nright to N
Assign Sleft and Sright to them, resp.
Repeat procedure for N = Nleft and N = Nright
8

Decision tree > CART > Building > Purity


(Im)Purity
Quality measure applied to each subset Sleft and Sright
Combination of the measures (e.g., weighted average)

Examples
Gini index =
Entropy =
Misclassification error =

Decision tree > CART > Properties


CART kNN

SVM

Intrinsically multiclass
Handles Apple and Orange features
Robustness to outliers
Works w/ "small" learning set
Scalability (large learning set)
Prediction accuracy
Parameter tuning
10

Outline
Machine learning
Decision tree
Random forest
Bagging
Random decision trees

Kernel-Induced Random Forest (KIRF)


Byproducts
Out-of-bag error
Variable importance
11

Random forest
Definition
Collection of unpruned CARTs
Rule to combine individual tree decisions

Purpose
Improve prediction accuracy

Principle
Encouraging diversity among the tree

Solution: randomness
Bagging
Random decision trees (rCART)

12

Random forest > Bagging


Bagging: Bootstrap aggregation
Technique of ensemble learning...
... to avoid over-fitting
Important since trees are unpruned
... to improve stability and accuracy

Two steps
Bootstrap sample set
Aggregation

13

Random forest > Bagging > Bootstrap


L: original learning set composed of p samples
Generate K learning sets Lk...
... composed of q samples, q p,...
... obtained by uniform sampling with replacement from L
In consequences, Lk may contain repeated samples

Random forest: q = p
Asymptotic proportion of unique samples in Lk = 100 (1 - 1/e) ~ 63%
The remaining samples can be used for testing

14

Random forest > Bagging > Aggregation


Learning
For each Lk, one classifier Ck (rCART) is learned

Prediction
S: a new sample
Aggregation = majority vote among the K predictions/votes Ck(S)

15

Random forest > Random decision tree


All labeled samples initially assigned to root node
Nroot node
With node N do

Find the feature F among a random subset of features + threshold value T...
... that split the samples assigned to N into 2 subsets Sleft and Sright...
... so as to maximize the label purity within these subsets

Assign (F,T) to N

If Sleft

else

and Sright too small to be splitted


Attach child leaf nodes Lleft and Lright to N
Tag the leaves with the most present label in Sleft and Sright, resp.
Attach child nodes Nleft and Nright to N
Assign Sleft and Sright to them, resp.
Repeat procedure for N = Nleft and N = Nright

Random subset of features

Random drawing repeated at each node


For D-dimensional samples, typical subset size = round(sqrt(D)) (also round(log2(x)))
Increases diversity among the rCARTs + reduces computational load

Typical purity: Gini index


16

Random forest > Properties


RF

CART kNN

SVM

Intrinsically multiclass
Handles Apple and Orange features
Robustness to outliers
Works w/ "small" learning set
Scalability (large learning set)
Prediction accuracy
Parameter tuning
17

Random forest > Illustration

1 rCART

10 rCARTs

100 rCARTs

500 rCARTs
18

Random forest > Limitations


Oblique/curved frontiers
Staircase effect
Many pieces of hyperplanes

Fundamentally discrete
Functional data? (Example: curves)

19

Outline
Machine learning
Decision tree
Random forest
Bagging
Random decision trees

Kernel-Induced Random Forest (KIRF)


Byproducts
Out-of-bag error
Variable importance
20

Kernel-Induced Random Forest (KIRF)


Random forest
Sample S is a vector
Features of S = components of S

Kernel-induced features
Learning set L = { Si, i [1..N] }
Kernel K(x,y)
Features of sample S = { Ki(S) = K(Si, S), i [1..N] }
Samples S and Si can be vectors or functional data

21

Kernel > Kernel trick


Kernel trick
Maps samples into an inner product space...
... usually of higher dimension (possibly infinite)...
... in which classification (or regression) is easier
Typically linear

Kernel K(x,y)
Symmetric
Positive semi-definite (Mercer's condition):


Note: mapping needs not to be known (might not even have an
explicit representation; e.g., Gaussian kernel)
22

Kernel > Examples


Polynomial (homogeneous):
Polynomial (inhomogeneous):
Hyperbolic tangent:
Gaussian:
Function of the distance between samples
Straightforward application to functional data of a metric space
E.g., curves

23

KIRF > Illustration


Gaussian kernel
Some similarity with vantage-point tree

Reminder: RF w/ 100 rCARTs


KIRF w/ 100 rCARTs

24

KIRF > Limitations


Which kernel?
Which kernel parameters?

No orange and apple handling anymore


(xy or (x - y)2)

Computational load (kernel evaluations)


Especially during learning

Needs to store samples


(Instead of feature indices in Random forest)

25

Outline
Machine learning
Decision tree
Random forest
Bagging
Random decision trees

Kernel-Induced Random Forest (KIRF)


Byproducts
Out-of-bag error
Variable importance
26

Byproduct > Reminder


To grow one rCART
Bootstrap sample set from learning set L
Remaining samples
Called out-of-bag samples
Can be used for testing

Two points of view


For one rCART, out-of-bag samples = L \ Bootstrap samples
Used for variable importance
For one sample S of L, set of rCARTs for which S was out-of-bag
Used for out-of-bag error

27

Byproduct > Out-of-bag error


For each sample S of the learning set

Look for all the rCARTs for which S was out-of-bag


Build the corresponding sub-forest
Predict the class of S with it
Error = is prediction correct?

Out-of-bag error = average over all samples of S


Note: predictions not made using the whole forest...
... but with some aggregation

Provides an estimation of the generalization error


Can be used to decide when to stop adding trees to the forest

28

Byproduct > Variable importance


For each rCART
Compute out-of-bag error OOBoriginal
Fraction of misclassified out-of-bag samples
Consider the ith feature/variable of the samples
Randomly permute its values among the out-of-bag samples
Re-compute out-of-bag error OOBpermutation
rCART-level importance(i) = OOBpermutation - OOBoriginal

Variable importance(i) = average over all rCARTs


Note: rCART-based errors (no aggregation)
Avoid attenuation of individual errors

29

An introduction
to random forests
Thank you for your attention

You might also like