05 Classification And Prediction

Objectives
Introduction
What is Classification?
Classification vs Prediction
Supervised and Unsupervised Learning
Data Preparation
D t P ti
Classification Accuracy
Classification and Prediction
Lecture 5/DMBI/IKI83403T/MTI/UI ID3 Algorithm
Yudho Giri Sucahyo, Ph.D, CISA (yudho@cs.ui.ac.id)
Information Gain
Faculty of Computer Science, University of Indonesia Bayesian Classification
Predictive Modelling
2 University of Indonesia

Introduction What is Classification? – A two step process
two-step
Databases are rich with hidden information that can be used Model construction:
for making intelligent business decisions. Each tuple is assumed to belong to a predefined class, as
Classification and prediction can be used to extract models determined by one of the attributes, called the class label.
describing i
d ibi important d classes or to predict f
data l di future data
d Data tuples are also referred to as samples, examples, or objects.
trends.
All tuples used for construction is called training set.
Classification predicts categorical labels. Ex: categorize bank
labels
loan applications safe or risky. Since the class label of each training sample is provided
Prediction models continuous-valued functions. Ex: predict the supervised learning. In clustering (unsupervised learning),
expenditures of potential customers on computer equipment the l labels f
th class l b l of each training sample i not known, and th
ht i i l is t k d the
given their income and occupation. number or set of classes to be learned may not be known in
Typical Applications: advance.
Credit approval, target marketing, The model is represented in the following forms:
Medical diagnosis, treatment effectiveness analysis
g , y Classification rules, (
, (IF-THEN statements), decision tree, mathematical
), ,
formulae
University of Indonesia

What is Classification? – A two step process (2)
two-step Classification Process (1)
The model is used for classifying future or Classification
unknown objects. Algorithms
Training
First, the predictive accuracy of the model is estimated Data
The known label of test sample is compared with the classified result
from the model.
Accuracy rate is the percentage of test set samples that are correctly
classified by the model. NAM E RANK YEARS TENURED Classifier
(Model)
Test set is independent of training set otherwise over-fitting (it may M ike Assistant Prof 3 no
have incorporated some particular anomalies of the training data that M ary Assistant Prof 7 yes
are not present in the overall sample population) will occur. Bill Professor 2 yes
If the accuracy of the model is considered acceptable the
acceptable, Jim Associate Prof 7 yes
model can be used to classify future objects for which the Dave Assistant Prof 6 no
IF rank = ‘professor’
OR years > 6
class label is not known (unknown, p
( previously unseen data).
y ) Anne Associate Prof 3 no THEN tenured = ‘yes’
yes
6
5 University of Indonesia University of Indonesia

Classification Process (2) What is Prediction?
Prediction is similar to classification
Classifier
First, construct model.
Second, use model to predict future or unknown objects
p j
Major method for prediction is regression:
Testing Linear and multiple regression
Data Unseen Data
Non-liner regression

(Jeff, Professor, 4) Prediction is different from classification
Classification refers to predict categorical class label.
NAM E RANK YEARS TENURED
Tom Assistant Prof 2 no Tenured? Prediction refers to predict continuous value.
M erlisa Associate Prof 7 no
George Professor 5 yes
Joseph Assistant Prof 7 yes
es
7

University of Indonesia

Classification vs Prediction Supervised vs Unsupervised Learning
Sending out promotional literature to every new Supervised learning (classification)
customer in the database can be quite costly. A more cos- Supervision: The training data (observations, measurements,
efficient method would be to target only those new etc.) are accompanied by labels indicating the class of the
) p y g
customers who are likely to purchase a new computer
h lik l h observations
classification.
Based on the training set to classify new data
g y
Predict h
P d the number of major purchases that a customer
b f h h
will make during a fiscal year prediction. Unsupervised learning (clustering)
We are given a set of measurements, observations etc with
measurements observations,
the aim of establishing the existence of classes or clusters in
the data
No training data, or the “training data” are not accompanied
by class labels

Issues – Data Preparation Issues – Data Preparation
Data preprocessing can be used to help improve the Data Transformation
accuracy, efficiency, and scalability of the classification or Data can be generalized to higher-level concepts.
prediction process. Useful fot continuous-valued attributes.
Data Cleaning Income can be generalized low, medium, high.
Remove/reduce noise and the treatment of missing values Street city.
Relevance Analysis Generalization compresses the original training data, fewer
Many of the attributes in the data may be irrelevant to the input/output operations may be involved during learning.
classification or prediction task. Ex: data recording the day of When i
Wh using neural networks (or other methods involving
l t k ( th th d i l i
the week on which a bank loan application was filed is unlikely distance measurements), data may also be normalized.
to be relevant to the success of the application
application.
Other attributes may be redundant.
This step is known as feature selection
selection.

Comparing Classification Method Classification Accuracy: Estimating Error Rates
Predictive accuracy Partition: Training-and-testing
Speed and scalability use two independent data sets, e.g., training set (2/3), test
time to construct the model
set(1/3)
time to use the model
used for data set with large number of samples
Robustness
handling noise and missing values Cross-validation
Scalability divide the data set into k subsamples
efficiency in large databases (not memory resident data)
use k-1 subsamples as training data and one sub-sample as test
p g p
Interpretability: data --- k-fold cross-validation
the level of understanding and insight provided by the model
for data set with moderate size
Goodness of rules
decision tree size Bootstrapping (leave-one-out)
the compactness of classification rules for small size data

What is a decision tree? Training Dataset
A decision tree is a flow-chart-like tree structure. An Example
Outlook Tempreature Humidity Windy Class
sunny hot high false N
Internal node denotes a test on an attribute from Quinlan’s sunny hot high true N
Branch represents an outcome of the test ID3 overcast hot high
g false P
All tuples in branch have the same value for the tested rain mild high false P
attribute. rain cool normal false P
Leaf node represents class label or class label distribution. rain cool normal tr e
true N
overcast cool normal true P
To classify an unknown sample, the attribute values of the sunnyy mild high
g false N
sample are tested against the decision tree. A path is traced sunny cool normal false P
from the root to a leaf node that holds the class prediction rain mild normal false P
for h
f that sample. l sunny mild
ild normal t
l true P
overcast mild high true P
Decision trees can easily be converted to classification rules. overcast hot normal false P
rain mild high true N

A Sample Decision Tree Decision-Tree
Decision Tree Classification Methods
The basic top-down decision tree generation approach
Outlook
usually consists of two phases:

sunny
Tree construction
overcast
overcast rain
At start, all the training examples are at the root.
Partition examples recursively based on selected
humidity P windy
attributes.
Tree pruning
high normal true false
Aiming at removing tree branches that may lead to errors
when classifying t t d t (t i i d t may contain noise,
h l if i test data (training data t i i
N P N P
outliers, …)


Choosing Split Attribute –
ID3 Algorithm Information Gain (ID3/C4 5) (1)
(ID3/C4.5)
All attributes are categorical Assume all attributes to be categorical (discrete-values).
Create a node N; Continuous-valued attributes must be discretized.
if samples are all of the same class C, then
Used to select the test attribute at each node in the tree.
return N as a leaf node labeled with C
if attribute-list is empty then Also called measure of the goodness of split.
return N as a leaf node labeled with the most common class The attribute with the highest information gain is chosen
g g
select split-attribute with highest information gain as the test attribute for the current node.
label N with the split-attribute
for
f each value Ai of split-attribute, grow a branch from Node N
h l f lit tt ib t b hf N d
let Si be the branch in which all tuples have the value Ai for split- attribute
if Si is empty then
attach a leaf labeled with the most common class
Else recursively run the algorithm at Node Si
until all branches reach leaf nodes


Information Gain (ID3/C4 5) (2)
(ID3/C4.5) Information Gain (ID3/C4 5) (3)
(ID3/C4.5)
Assume that there are two classes, P and N. The attribute A is selected such that the information gain
Let h
L the set of examples S contain p elements of class P and n
f l l f l d gain(A) = I(p, n) - E(A)
elements of class N.
The amount of information needed to decide if an arbitrary
information, is maximal, that is, E(A) is minimal since I(p, n) is the same to all
, , ( ) (p, )
example in S belong to P or N is defined as attributes at a node.
p p n n In the given sample data, attribute outlook is chosen to split at
data
I ( p, n ) = − lo g 2 − lo g 2
p+n p+n p+n p+n
the root :
Assume that using attribute A as the root in the tree will partition
S in sets {S1, S2 , …, Sv}
}. gain(outlook) = 0 246
i ( tl k) 0.246
If Si contains pi examples of P and ni examples of N, the information gain(temperature) = 0.029
needed to classify objects in all subtrees Si :
gain(humidity) = 0.151
v pi + ni gain(windy) = 0.048
E( A) = ∑ I ( pi , ni)
i =1 p+n

Information Gain (ID3/C4 5) (3)
(ID3/C4.5) Information Gain (ID3/C4 5) (4)
(ID3/C4.5)
Examples: Next, compute the entropy of each attribute. Let’s start with the
See Table 7.1. attribute age.
ib
For age = “<= 30”: s11 = 2 s21 = 3 I (s11, s21) = 0.971
Class label: buys_computer. Two values:YES, NO.
For age = “31 40”: s12 = 4 s22 = 0 I (s12, s22) = 0
31..40 :
m = 2. C1 correspond to yes, C2 correspond to no.
For age = “>40”: s13 = 3 s23 = 2 I (s13, s23) = 0.971
9 samples of class yes and 5 samples of class no.
Using equation (7.2), the expected information needed to classify
(7 2)
Compute the expected information needed to classify a given
a given sample if the samples are partitioned by age is
sample 5 4 5
E ( age ) = I ( s 11 , s 21 ) + I ( s 12 , s 22 ) + I ( s 13 , s 23 ) = 0 . 694
9 9 5 5 14 14 14
I ( s1 , s 2 ) = I (9,5) = − log − log = 0 .940 Hence, the gain in information from such a partitioning:
14 2 14 14 2 14

Gain(age) = I (s1, s2) – E (age) = 0.246
( g ) ( ( g )
Similarly, we can compute Gain(income) = 0.029, Gain(student) =
0.151, Gain(Credit_rating) = 0.048.


How to use a tree? Tree Pruning
Directly A decision tree constructed using the training data may have
test the attribute value of unknown sample against the tree. too many branches/leaf nodes.
/ f
A path is traced from root to a leaf which holds the label Caused by noise, overfitting
Indirectly May
M result poor accuracy for unseen samples
l f l
decision tree is converted to classification rules Prune the tree: merge a subtree into a leaf node.
one rule is created for each path from the root to a leaf Using
U i a set of d different from the training data.
f data diff f h i i d
IF-THEN is easier for humans to understand At a tree node, if the accuracy without splitting is higher than
Example:
the accuracy with splitting replace the subtree with a leaf node
splitting, node,
label it using the majority class.
IF age = “<=30” AND student = “no” THEN buys_computer = “no”
Pruning Criterion:
Pessimistic pruning: C4.5
MDL: SLIQ and SPRINT
Cost complexity pruning: CART

Classification and Databases Classifying Large Dataset
Classification is a classical problem extensively studied by Decision trees seem to be a good choice
statisticians relatively faster learning speed than other classification
AI, especially machine learning researchers methods
Database researchers re-examined the problem in the can be converted into simple and easy to understand
classification rules
context of large databases
can be used to generate SQL queries for accessing databases
most previous studies used small size data, and most
algorithms are memory resident has comparable classification accuracy with other methods
Recent data mining research contributes to Classifying data sets with millions of examples and a few
data-sets
Scalability hundred even thousands attributes with reasonable
Generalization-based classification
speed.
speed
Parallel and distributed processing


Scalable Decision Tree Methods Previous Efforts on Scalability
Most algorithms assume data can fit in memory. Incremental tree construction (Quinlan’86)
( )
Data mining research contributes to the scalability issue, using partial data to build a tree.
especially for decision trees.
p y testing other examples and those mis-classified ones are used
g p
to rebuild the tree interactively.
Successful examples
Data reduction (Cattlet’91)
SLIQ (EDBT’96 -- Mehta et al ’96)
(EDBT 96 al. 96)
reducing data size by sampling and discretization.
SPRINT (VLDB96 -- J. Shafer et al.’96)
still a main memory algorithm.
PUBLIC (VLDB98 -- Rastogi & Shim 98)
Shim’98)
Data partition and merge (Chan and Stolfo’91)
RainForest (VLDB98 -- Gehrke, et al.’98)
partitioning data and building trees for each partition.
merging multiple trees into a combined tree.
experiment results indicated reduced classification accuracy.


Presentation of Classification Rules Other Classification Methods

Bayesian Classification
Neural Networks
Genetic Algorithm
Rough Set Approach
k-Nearest Neighbor Classifier
Case-Based Reasoning (CBR)
Fuzzy Logic
Support Vector Machine (SVM)

Bayesian Classification Bayes Theorem (1)
Bayesian classifiers are statistical classifiers. Let X be a data sample whose class label is unknown.
They can predict class membership probabilities, such as Let H be some hypothesis, such as that the data sample X
the probability that a given sample belongs to a particular belongs to a specified class C.
class. We want to determine P(H|X), the probability the the
Bayesian classification is based on Bayes theorem. hypothesis H holds given the observed data sample X.
Naive Bayesian Classifier is comparable in performance P(H|X) is the posterior probability or a posteriori
with decision tree and neural network classifiers. probability, of H conditioned on X.
Bayesian classifiers also have high accuracy and speed Support the world of data samples consists of fruits, described
when applied to large databases. by their color and shape., Suppose that X is red and round, and
that H i the h
h is h hypothesis that X i an apple. Th P(H|X)
h i h is l Then
reflects our confidence that X is an apple given that we have
seen that X is red and round.

Bayes Theorem (2) Predictive Modeling in Databases
P(H) is the prior probability or a priori probability, of H. What if we would like to predict a continuous value,
The probability that any given data sample is an apple, rather than a categorical label?
regardless of how the data sample looks. Prediction of continuous values can be modeled by statistical
The posterior probability is based on more information techniques of regression.
h i f i
(such as background knowledge) than the prior Example:
probability which i i d
b bili hi h is independent of X
d f X. A m dle t predict the salary of c lle e graduates with 10 years of
modle to redict salar f college rad ates ith ears f
work experience.
Bayes theorem is P(H | X ) = P( X | H )P(H ) Potential sales of a new product given its price.
P( X )
(X
Many problems can be solved by linear regression.
See example 7.4 for example on Naive Bayesian Software packages for solving regression problems:
Classification. SAS, SPSS, S-Plus


Linear Regression Prediction: Numerical Data
Data are modeled using a straight line.
The simplest form of regression
Bivariate liner regressions models a random variable Y
g
(called a response variable), as a linear function of another
random variable, X (called a predictor variable)
Y=α+βX
See Example 7.6 for an example of linear regression.
p p g
Other regression models
Multiple regression
Log-linear models


Prediction: Categorical Data Conclusion
Classification is an extensively studied problem (mainly in
statistics, machine learning & neural networks)
Classification is probably one of the most widely used data
mining techniques with a lot of applications.
Scalability is still an important issue for database applications.
Combining classification with database techniques should be a
g q
promising research topic.
Research Direction: Classification of non relational data e g
non-relational data, e.g.,
text, spatial, multimedia, etc..


References References (2)
C. Apte and S. Weiss. Data mining with decision trees and decision rules. Future Generation J. Magidson. The chaid approach to segmentation modeling: Chi-squared automatic
interaction detection. In R. P. Bagozzi, editor, Advanced Methods of Marketing Research, p g
g , , g , pages
Computer S t
C t Systems, 13, 1997.
13 1997
118-159. Blackwell Business, Cambridge Massechusetts, 1994.
L. Breiman, J. Friedman, R. Olshen, and C. Stone. Classification and Regression Trees. Wadsworth M. Mehta, R. Agrawal, and J. Rissanen. SLIQ : A fast scalable classifier for data mining. In Proc.
International Group, 1984.
p 1996 Int. Conf. Extending Database Technology (EDBT'96), Avignon, France, March 1996.

P. K. Chan and S. J. Stolfo. Learning arbiter and combiner trees from partitioned data for scaling S. K. Murthy, Automatic Construction of Decision Trees from Data: A Multi-Diciplinary Survey,
Data Mining and Knowledge Discovery 2(4): 345-389, 1998
machine learning. In Proc. 1st Int. Conf. Knowledge Discovery and Data Mining (KDD'95), pages
J
J. R. Quinlan. Bagging, boosting, and c4.5. In Proc. 13th Natl. Conf. on Artificial Intelligence
gg g g g
39-44, M
39 44 Montreal, Canada, August 1995.
l C d A 1995 (AAAI'96), 725-730, Portland, OR, Aug. 1996.
U. M. Fayyad. Branching on attribute values in decision tree generation. In Proc. 1994 AAAI Conf., R. Rastogi and K. Shim. Public: A decision tree classifer that integrates building and pruning. In
p g
pages 601-606, AAAI Press, 1994.
, , Proc. 1998 Int. Conf. Very Large Data Bases, 404-415, New York, NY, August 1998.
J. Shafer, R. Agrawal, and M. Mehta. SPRINT : A scalable parallel classifier for data mining. In
J. Gehrke, R. Ramakrishnan, and V. Ganti. Rainforest: A framework for fast decision tree
Proc. 1996 Int. Conf. Very Large Data Bases, 544-555, Bombay, India, Sept. 1996.
construction of large datasets. In Proc. 1998 Int. Conf. Very Large Data Bases, pages 416-427, New
S. M. Weiss and C. A. Kulikowski. Computer Systems that Learn: Classification and Prediction
York, NY, August 1998. Methods from Statistics, Neural Nets, Machine Learning, and Expert Systems. Morgan
Kaufman, 1991.
M. Kamber, L. Winstone, W. Gong, S. Cheng, and J. Han. Generalization and decision tree induction:
Efficient classification in data mining. In Proc. 1997 Int. Workshop Research Issues on Data
Engineering (RIDE'97), pages 111-120, Birmingham, England, April 1997.

05 Classification And Prediction

More Related Content

What's hot (20)

Viewers also liked (11)

Similar to 05 Classification And Prediction (20)

More from Achmad Solichin (20)

Recently uploaded (20)

05 Classification And Prediction