Data Analytics

(CS40003)
Lecture #11
Sensitivity Analysis

Dr. Debasis Samanta

Associate Professor
Department of Computer Science & Engineering
 Topics Covered in this Presentation
 Introduction

 Estimation Strategies

 Accuracy Estimation

 Error Estimation

 Statistical Estimation

 Performance Estimation

 ROC Curve

CS 40003: Data Analytics 2

 Introduction
 A classifier is used to predict an outcome of a test data

 Such a prediction is useful in many applications

 Business forecasting, cause-and-effect analysis, etc.

 A number of classifiers have been evolved to support the activities.

 Each has their own merits and demerits

 There is a need to estimate the accuracy and performance of the classifier with
respect to few controlling parameters in data sensitivity

 As a task of sensitivity analysis, we have to focus on

 Estimation strategy
 Metrics for measuring accuracy
 Metrics for measuring performance
CS 40003: Data Analytics 3
 Estimation Strategy

CS 40003: Data Analytics 4

Planning for Estimation
 Using some “training data”, building a classifier based on certain principle is
called “learning a classifier”.

 After building a classifier and before using it for classification of unseen

instance, we have to validate it using some “test data”.

 Usually training data and test data are outsourced from a large pool of data
already available.
Learning
Training technique
data

Test data
split CLASSIFIER

CS 40003: Data Analytics

Data set Estimation 5
Estimation Strategies
 Accuracy and performance measurement should follow a strategy. As the
topic is important, many strategies have been advocated so far. Most
widely used strategies are

 Holdout method

 Random subsampling

 Cross-validation

 Bootstrap approach

CS 40003: Data Analytics 6

Holdout Method
 This is a basic concept of estimating a prediction.

 Given a dataset, it is partitioned into two disjoint sets called training set and
testing set.

 Classifier is learned based on the training set and get evaluated with testing set.

 Proportion of training and testing sets is at the discretion of analyst; typically

1:1 or 2:1, and there is a trade-off between these sizes of these two sets.

 If the training set is too large, then model may be good enough, but estimation
may be less reliable due to small testing set and vice-versa.

CS 40003: Data Analytics 7

Random Subsampling
 It is a variation of Holdout method to overcome the drawback of over-
presenting a class in one set thus under-presenting it in the other set and
vice-versa.

 In this method, Holdout method is repeated k times, and in each time, two
disjoint sets are chosen at random with a predefined sizes.

 Overall estimation is taken as the average of estimations obtained from

each iteration.

CS 40003: Data Analytics 8

Cross-Validation
 The main drawback of Random subsampling is, it does not have
control over the number of times each tuple is used for training and
testing.

 Cross-validation is proposed to overcome this problem.

 There are two variations in the cross-validation method.

 k-fold cross-validation

 N-fold cross-validation

CS 40003: Data Analytics 9

k-fold Cross-Validation
Dataset
   consisting of N tuples is divided into k (usually, 5 or 10) equal,
mutually exclusive parts or folds (, and if N is not divisible by k, then the last
part will have fewer tuples than other (k-1) parts.

 A series of k runs is carried out with this decomposition, and in ith iteration is
used as test data and other folds as training data
 Thus, each tuple is used same number of times for training and once for testing.

 Overall estimate is taken as the average of estimates obtained from each

iteration. D1 Fold 1

Learning
Di technique
Fold i

Data set
Dk

Fold k
CLASSIFIER

CS 40003: Data Analytics Accuracy Performance 10

N-fold Cross-Validation
In k-fold cross-validation method, part of the given data is used in training

 
with k-tests.

 N-fold cross-validation is an extreme case of k-fold cross validation, often

known as “Leave-one-out’’ cross-validation.

 Here, dataset is divided into as many folds as there are instances; thus, all
most each tuple forming a training set, building N classifiers.

 In this method, therefore, N classifiers are built from N-1 instances, and
each tuple is used to classify a single test instances.

 Test sets are mutually exclusive and effectively cover the entire set (in
sequence). This is as if trained by entire data as well as tested by entire data
set.

CS 40003: Data Analytics 11

 Overall estimation is then averaged out of the results of N classifiers.
 N-fold Cross-Validation : Issue
 So far the estimation of accuracy and performance of a classifier model is
concerned, the N-fold cross-validation is comparable to the others we have
just discussed.

 The drawback of N-fold cross validation strategy is that it is

computationally expensive, as here we have to repeat the run N times; this
is particularly true when data set is large.

 In practice, the method is extremely beneficial with very small data set
only, where as much data as possible to need to be used to train a classifier.

CS 40003: Data Analytics 12

Bootstrap Method
 The Bootstrap method is a variation of repeated version of Random
sampling method.

 The method suggests the sampling of training records with replacement.

 Each time a record is selected for training set, is put back into the original pool of
records, so that it is equally likely to be redrawn in the next run.

 In other words, the Bootstrap method samples the given data set uniformly
with replacement.

 The rational of having this strategy is that let some records be occur more
than once in the samples of both training as well as testing.

 What is the probability that a record will be selected more than once?
CS 40003: Data Analytics 13
Bootstrap Method
Suppose, we have given a data set of N records. The data set is sampled N times

 
with replacement, resulting in a bootstrap sample (i.e., training set) of I samples.
 Note that the entire runs are called a bootstrap sample in this method.

 There are certain chance (i.e., probability) that a particular tuple occurs one or
more times in the training set

 If they do not appear in the training set, then they will end up in the test set.

 Each tuple has a probability of being selected (and the probability of not being
selected is .
 We have to select N times, so the probability that a record will not be chosen during
the whole run is
 Thus, the probability that a record is chosen by a bootstrap sample is
 For a large value of N, it can be proved that
 record chosen in a bootstrap sample is = 0.632

CS 40003: Data Analytics 14

Bootstrap Method : Implication

 This is why, the Bootstrap method is also known as 0.632 bootstrap method

CS 40003: Data Analytics 15

 Accuracy Estimation

CS 40003: Data Analytics 16

Accuracy Estimation
We have learned how a classifier system can be tested. Next, we are to learn

 
the metrics with which a classifier should be estimated.

 There are mainly to things to be measured for a given classifier

 Accuracy
 Performance

 Accuracy estimation
 If N is the number of instances with which a classifier is tested and p is the number
of correctly classified instances, the accuracy can be denoted as

 Also, we can say the error rate (i.e., is misclassification rate) denoted by is denoted
by
CS 40003: Data Analytics 17
Accuracy : True and Predictive
Now, this accuracy may be true (or absolute) accuracy or predicted (or

 
optimistic) accuracy.

 True accuracy of a classifier is the accuracy when the classifier is tested

with all possible unseen instances in the given classification space.
 However, the number of possible unseen instances is potentially very large
(if it is not infinite)
 For example, classifying a hand-written character
 Hence, measuring the true accuracy beyond the dispute is impractical.

 Predictive accuracy of a classifier is an accuracy estimation for a given

test data (which are mutually exclusive with training data).
 If the predictive accuracy for test set is and if we test the classifier with a
different test set it is very likely that a different accuracy would be obtained.

 The predictive accuracy when estimated with a given test set it should be
acceptable without any objection
CS 40003: Data Analytics 18
Predictive Accuracy
Example 11.1 : Universality of predictive accuracy
 Consider a classifier model MD developed with a training set D using an
algorithm M.

 Two predictive accuracies when MD is estimated with two different training

sets T1 and T2 are
(MD)T1 = 95%
(MD)T2 = 70%

 Further, assume the size of T1 and T2 are

|T1| = 100 records
|T2| = 5000 records.

 Based on the above mentioned estimations, neither estimation is acceptable

beyond doubt.
CS 40003: Data Analytics 19
Predictive Accuracy
 With the above-mentioned issue in mind, researchers have proposed two
heuristic measures

 Error estimation using Loss Functions

 Statistical Estimation using Confidence Level

 In the next few slides, we will discus about the two estimations

CS 40003: Data Analytics 20

Error Estimation using Loss Functions
Let T be a matrix comprising with N test tuples

 
X1 y1
X2 y2

XN yN
N×(n+1)

where Xi (i = 1, 2, …, N) is the n-dimensional test tuples with associated outcome yi.

 Suppose, corresponding to (Xi, yi), classifier produces the result (Xi, )

 Also, assume that denotes a difference between and (following certain difference (or
similarity), (e.g., = 0, if there is a match else 1)

 The two loss functions measure the error between (the actual value) and (the predicted
value) are

Absolute error:
Squred error:
CS 40003: Data Analytics 21
Error Estimation using Loss Functions
Based

  on the two loss functions, the test error (rate) also called generalization
error, is defined as the average loss over the test set T. The following two
measures for test errors are
Mean Absolute Error (MAE):
Mean Squared Error(MSE): ):

 Note that, MSE aggregates the presence of outlier.

 In addition to the above, a relative error measurement is also known. In this measure,
the error is measured relative to the mean value calculated as the mean of yi (i = 1, 2,
…, N) of the training data say D. Two measures are
Relative Absolute Error (RAE:
Relative Squared Error (RSE):

CS 40003: Data Analytics 22

Statistical Estimation using Confidence Level
 In fact, if we know the value of predictive accuracy, say , then we can guess the true
 
accuracy within a certain range given a confidence level.

 Confidence level: The concept of “confidence level ” can be better understood with the
following two experiments, related to tossing a coin.

 Experiment 1: When a coin is tossed, there is a probability that the head will occur. We have
to experiment the value for this probability value. A simple experiment is that the coin is
tossed many times and both numbers of heads and tails are recorded.

N=10 N=50 N=100 N=250 N=500 N=1000

H T H T H T H T H T H T

3 7 29 21 54 46 135 115 241 259 490 510

0.30 0.70 0.58 0.42 0.54 0.46 0.54 0.46 0.48 0.42 0.49 0.51

 Thus, we can say that after a large number of trials in each experiment.

CS 40003: Data Analytics 23

Statistical Estimation using Confidence Level
Experiment
   2: A similar experiment but with different counting is
conducted to learn the probability that a coin is flipped its head 20 times
out of 50 trials. This experiment is popularly known as Bernoulli's trials.
It can be stated as follows.

P
 where N = Number of trials
 v = Number of outcomes that an event occurs.
 p = Probability that the event occur

 Thus, if p = 0.5, then

 Note:
 Also, we may note the following
 Mean = N×p = 50×0.5 = 25 and Variance = p× (1-p) ×N = 50×0.5×0.5 = 12.5

CS 40003: Data Analytics 24

Statistical Estimation using Confidence Level
The
   task of predicting the class labels of test records can also be
considered as a binomial experiment, which can be understood as
follows. Let us consider the following.
 N = Number of records in the test set.
 n = Number of records predicted correctly by the classifier.
 = n/N, the observed accuracy (it is also called the empirical accuracy).
 = the true accuracy.

 Let and denotes the lower and upper bound of a confidence level . Then
the confidence interval for is given by
~
  ∈− ∈
(
𝑃 𝜏 ≤
𝐿
∝
√ ∈ ( 1 −∈ ) / 𝑁
𝑈
≤ 𝜏 ∝ =𝛼
)
 If is the mean of and , then we can write
 ~
∈ =∈ ± 𝜏 𝛼 × √ ∈ (1 − ∈ )/ 𝑁
CS 40003: Data Analytics 25
Statistical Estimation using Confidence Level
~
∈=∈± 𝝉 𝜶 × √∈(𝟏−∈)/𝑵
 

A  table of with different values of can be obtained from any book on

statistics. A small part of the same is given below.

0.5 0.7 0.8 0.9 0.95 0.98 0.99

0.67 1.04 1.28 1.65 1.96 2.33 2.58

 Thus, given a confidence level , we shall be able to know the value of and
hence the true accuracy (), if we have the value of the observed accuracy ().

 Thus, knowing a test data set of size N, it is possible to estimate the true
accuracy!

CS 40003: Data Analytics 26

Statistical Estimation using Confidence Level
Example
  11.2: True accuracy from observed accuracy
A classifier is tested with a test set of size 100. Classifier predicts 80 test tuples
correctly. We are to calculate the following.
a) Observed accuracy
b) Mean error rate
c) Standard error
d) True accuracy with confidence level 0.95.
Solution:
e) = 80/100 = 0.80 So error (p) = 0.2
f) Mean error rate = p×N = 0.2×N = 20
g) Standard error rate (σ) = = = 0.04
h) = 0.8±0.04×1.96 = 0.7216 with =1.96 and = 0.95.

CS 40003: Data Analytics 27

Statistical Estimation using Confidence Level

Note:
 
 Suppose, a classifier is tested k times with k different test sets. If i denotes the
predicted accuracy when tested with test set Ni in the i-th run (1≤ i ≤ k), then
the overall predicted accuracy is

Thus, is the weighted average of values. The standard error and true accuracy
at a confidence are

CS 40003: Data Analytics 28

 Performance Estimation

CS 40003: Data Analytics 29

Performance Estimation of a Classifier
 Predictive accuracy works fine, when the classes are balanced
 That is, every class in the data set are equally important

 In fact, data sets with imbalanced class distributions are quite common in
many real life applications

 When the classifier classified a test data set with imbalanced class
distributions then, predictive accuracy on its own is not a reliable indicator of
a classifier’s effectiveness.

Example 11.3: Effectiveness of Predictive Accuracy

 Given a data set of stock markets, we are to classify them as “good” and “worst”.
Suppose, in the data set, out of 100 entries, 98 belong to “good” class and only 2
are in “worst” class.

 With this data set, if classifier’s predictive accuracy is 0.98, a very high value!
 Here, there is a high chance that 2 “worst” stock markets may incorrectly be classified as “good”
CS 40003: Data Analytics 30
Performance Estimation of a Classifier
 Thus, when the classifier classified a test data set with imbalanced class
distributions, then predictive accuracy on its own is not a reliable indicator of
a classifier’s effectiveness.

 This necessitates an alternative metrics to judge the classifier.

 Before exploring them, we introduce the concept of Confusion matrix.

CS 40003: Data Analytics 31

Confusion Matrix
 A confusion matrix for a two classes (+, -) is shown below.

 There are four quadrants in the confusion matrix, which are symbolized as
below.
 True Positive (TP: f++) : The number of instances that were positive (+) and
correctly classified as positive (+v).

 False Negative (FN: f+-): The number of instances that were positive (+) and
incorrectly classified as negative (-). It is also known as Type 2 Error.

 False Positive (FP: f-+): The number of instances that were negative (-) and
incorrectly classified as (+). This also known as Type 1 Error.
CS 40003: Data Analytics 32
Confusion Matrix
Note:
 Np = TP (f++) + FN (f+-)
= is the total number of positive instances.

 Nn = FP(f-+) + Tn(f--)
= is the total number of negative instances.

 N = Np + Nn
= is the total number of instances.

 (TP + TN) denotes the number of correct classification

 (FP + FN) denotes the number of errors in classification.

 For a perfect classifier FP = FN = 0, that is, there would be no Type 1 or Type

2 errors.
CS 40003: Data Analytics 33
Confusion Matrix
Example 11.4: Confusion matrix
A classifier is built on a dataset regarding Good and Worst classes of stock markets.
The model is then tested with a test set of 10000 unseen instances. The result is shown
in the form of a confusion matrix. The result is self explanatory.

Class Good Worst Total Rate(%)

Good 6954 46 7000 99.34
Worst 412 2588 3000 86.27
Total 7366 2634 10000 95.52

Predictive accuracy?

CS 40003: Data Analytics 34

Confusion Matrix for Multiclass Classifier
 Having m classes, confusion matrix is a table of size m×m , where,
element at (i, j) indicates the number of instances of class i but
classified as class j.

 To have good accuracy for a classifier, ideally most diagonal entries

should have large values with the rest of entries being close to zero.

 Confusion matrix may have additional rows or columns to provide total

or recognition rates per class.

CS 40003: Data Analytics 35

Confusion Matrix for Multiclass Classifier
Example 11.5: Confusion matrix with multiple class

Following table shows the confusion matrix of a classification problem with six
classes labeled as C1, C2, C3, C4, C5 and C6.
Class C1 C2 C3 C4 C5 C6
C1 52 10 7 0 0 1
C2 15 50 6 2 1 2
C3 5 6 6 0 0 0
C4 0 2 0 10 0 1
C5 0 1 0 0 7 1
C6 1 3 0 1 0 24

Predictive accuracy?

CS 40003: Data Analytics 36

Confusion Matrix for Multiclass Classifier
 In case of multiclass classification, sometimes one class is important enough
to be regarded as positive with all other classes combined togather as negative.

 Thus a large confusion matrix of m*m can be concised into 2*2 matrix.

Example 11.6: m×m CM to 2×2 CM

 For example, the CM shown in Example 11.5 is transformed into a CM of size 2×2
considering the class C1 as the positive class and classes C2, C3, C4, C5 and C6
Class as negative.
combined together + -
+ 52 18
- 21 123

How we can calculate the predictive accuracy of the classifier model in this case?
Are the predictive accuracy same in both Example 11.5 and Example 11.6?
CS 40003: Data Analytics 37
Performance Evaluation Metrics
We now define a number of metrics for the measurement of a classifier.

 
 In our discussion, we shall make the assumptions that there are only two classes: + (positive)
and – (negative)

 Nevertheless, the metrics can easily be extended to multi-class classifiers (with some
modifications)

 True Positive Rate (TPR): It is defined as the fraction of the positive examples
predicted correctly by the classifier.

 This metrics is also known as Recall, Sensitivity or Hit rate.

 False Positive Rate (FPR): It is defined as the fraction of negative examples classified as
positive class by the classifier.

 This metric is also known as False Alarm Rate.

CS 40003: Data Analytics 38
Performance Evaluation Metrics
False
   Negative Rate (FNR): It is defined as the fraction of positive examples
classified as a negative class by the classifier.

 True Negative Rate (TNR): It is defined as the fraction of negative

examples classified correctly by the classifier

 This metric is also known as Specificity.

CS 40003: Data Analytics 39

 Performance Evaluation Metrics
  Positive Predictive Value (PPV): It is defined as the fraction of the positive

examples classified as positive that are really positive

It is also known as Precision.

F1 Score (F1): Recall (r) and Precision (p) are two widely used metrics employed
in analysis, where detection of one of the classes is considered more significant
than the others.
 It is defined in terms of (r or TPR) and (p or PPV) as follows.

Note
F1 represents the harmonic mean between recall and precision

High value of F1 score ensures that both Precision and Recall are reasonably high.
CS 40003: Data Analytics 40
Performance Evaluation Metrics
More

  generally, score can be used to determine the trade-off between Recall
and Precision as

 Both, Precision and Recall are special cases of when and , respectively.

CS 40003: Data Analytics 41

Performance Evaluation Metrics
A  more general metric that captures Recall, Precision as well as is defined in

the following.

Metric
Recall 1 1 0 1
Precision 1 0 1 0
+1 1 0

Note
 In fact, given TPR, FPR, p and r, we can derive all others measures.

 That is, these are the universal metrics.

CS 40003: Data Analytics 42

Predictive Accuracy (ε)
It  is defined as the fraction of the number of examples that are correctly

classified by the classifier to the total number of instances.

 This accuracy is equivalent to F with w = w = w = w =1.

w 1 2 3 4

CS 40003: Data Analytics 43

 Error Rate ()

The
  error rate is defined as the fraction of the examples that are
incorrectly classified.

Note
.

CS 40003: Data Analytics 44

Accuracy, Sensitivity and Specificity
Predictive
   accuracy () can be expressed in terms of sensitivity and specificity.

 We can write

Thus,

CS 40003: Data Analytics 45

Analysis with Performance Measurement Metrics
Based
   on the various performance metrics, we can characterize a classifier.

 We do it in terms of TPR, FPR, Precision and Recall and Accuracy

 Case 1: Perfect Classifier

When every instance is correctly classified, it is called the perfect classifier. In

this case, TP = P, TN = N and CM is
    Predicted Class

TPR = =1    
+ -

FPR = =0 + P 0
Actual
class

Precision = = 1 - 0 N

F1 Score = = 1
Accuracy = =1

CS 40003: Data Analytics 46

Analysis with Performance Measurement Metrics
   2: Worst Classifier
 Case

When every instance is wrongly classified, it is called the worst classifier. In this
case, TP = 0, TN = 0 and the CM is
    Predicted Class
TPR = =0    
+ -
FPR = = 1
+ 0 P
Precision = = 0

Actual
class
F1 Score = Not applicable - N 0

as Recall + Precision = 0
Accuracy = =0

CS 40003: Data Analytics 47

Analysis with Performance Measurement Metrics
   3: Ultra-Liberal Classifier
 Case

The classifier always predicts the + class correctly. Here, the False Negative
(FN) and True Negative (TN) are zero. The CM is
    Predicted Class
TPR = = 1    
+ -
FPR = = 1
+ P 0
Precision =

Actual
class
F1 Score = - N 0

Accuracy = =0

CS 40003: Data Analytics 48

Analysis with Performance Measurement Metrics
   4: Ultra-Conservative Classifier
 Case

This classifier always predicts the - class correctly. Here, the False Negative
(FN) and True Negative (TN) are zero. The CM is
    Predicted Class
TPR = = 0    
+ -
FPR = = 0
+ 0 p
Precision =

Actual
class
(as TP + FP = 0) - 0 N

F1 Score =
Accuracy = =0

CS 40003: Data Analytics 49

Predictive Accuracy versus TPR and FPR
 One strength of characterizing a classifier by its TPR and FPR is that they do
not depend on the relative size of P and N.

 The same is also applicable for FNR and TNR and others measures from CM.

 In contrast, the Predictive Accuracy, Precision, Error Rate, F1 Score, etc. are
affected by the relative size of P and N.

 FPR, TPR, FNR and TNR are calculated from the different rows of the CM.

 On the other hand Predictive Accuracy, etc. are derived from the values in both
rows.

 This suggests that FPR, TPR, FNR and TNR are more effective than
Predictive Accuracy, etc.

CS 40003: Data Analytics 50

 ROC Curves

CS 40003: Data Analytics 51

ROC Curves
 ROC is an abbreviation of Receiver Operating Characteristic come from the
signal detection theory, developed during World War 2 for analysis of radar
images.

 In the context of classifier, ROC plot is a useful tool to study the behaviour of
a classifier or comparing two or more classifiers.

 A ROC plot is a two-dimensional graph, where, X-axis represents FP rate

(FPR) and Y-axis represents TP rate (TPR).

 Since, the values of FPR and TPR varies from 0 to 1 both inclusive, the two
axes thus from 0 to 1 only.

 Each point (x, y) on the plot indicating that the FPR has value x and the TPR
value y.

CS 40003: Data Analytics 52

ROC Plot
 A typical look of ROC plot with few points in it is shown in the following
figure.

 Note the four cornered points are the four extreme cases of classifiers

Identify the four extreme classifiers.

CS 40003: Data Analytics 53

Interpretation of Different Points in ROC Plot
 Le us interpret the different points in the ROC plot.

 The four points (A, B, C, and D)

 A: TPR = 1, FPR = 0, the ideal model, i.e., the perfect classifier, no false results

 B: TPR = 0, FPR = 1, the worst classifier, not able to predict a single instance

 C: TPR = 0, FPR = 0, the model predicts every instance to be a Negative class, i.e., it is an ultra-conservative
classifier

 D: TPR = 1, FPR = 1, the model predicts every instance to be a Positive class, i.e., it is an ultra-liberal classifier

CS 40003: Data Analytics 54

Interpretation of Different Points in ROC Plot
 Le us interpret the different points in the ROC plot.

 The points on diagonals

 The diagonal line joining point C(0,0) and D(1,1) corresponds to random guessing

 Random guessing means that a record is classified as positive (0r negative) with a certain probability
 Suppose, a test set contacting N+ positive and N- negative instances. Suppose, the classifier guesses any instances with
probability p
 Thus, the random classifier is expected to correctly classify p.N+ of the positive instances and p.N- of the negative instances
 Hence, TPR = FPR = p
 Since TPR = FPR, the random classifier results reside on the main diagonals
CS 40003: Data Analytics 55
Interpretation of Different Points in ROC Plot
 Let us interpret the different points in the ROC plot.

 The points on the upper diagonal region

 All points, which reside on upper-diagonal region are corresponding to classifiers “good” as their
TPR is as good as FPR (i.e., FPRs are lower than TPRs)

 Here, X is better than Z as X has higher TPR and lower FPR than Z.

 If we compare X and Y, neither classifier is superior to the other

CS 40003: Data Analytics 56

Interpretation of Different Points in ROC Plot
 Let us interpret the different points in the ROC plot.

 The points on the lower diagonal region

 The Lower-diagonal triangle corresponds to the classifiers that are worst than random classifiers

 Note: A classifier that is worst than random guessing, simply by reversing its prediction, we can
get good results.

 W’(0.2, 0.4) is the better version than W(0.4, 0.2), W’ is a mirror reflection of W
CS 40003: Data Analytics 57
Tuning a Classifier through ROC Plot
 Using ROC plot, we can compare two or more classifiers by their TPR and
FPR values and this plot also depicts the trade-off between TPR and FPR of a
classifier.

 Examining ROC curves can give insights into the best way of tuning
parameters of classifier.

 For example, in the curve C2, the result is degraded after the point P.
Similarly for the observation C1, beyond Q the settings are not acceptable.
CS 40003: Data Analytics 58
Comparing Classifiers trough ROC Plot
 Two curves C1 and C2 are corresponding to the experiments to choose two
classifiers with their parameters.

 Here, C1 is better than C2 when FPR is less than 0.3.

 However, C2 is better, when FPR is greater than 0.3.

 Clearly, neither of these two classifiers dominates the other.

CS 40003: Data Analytics 59

Comparing Classifiers trough ROC Plot
 We can use the concept of “area under curve” (AUC) as a better method to compare two
or more classifiers.

 If a model is perfect, then its AUC = 1.

 If a model simply performs random guessing, then its AUC = 0.5

 A model that is strictly better than other, would have a larger value of AUC than the other.

 Here, C3 is best, and C2 is better than C1 as AUC(C3)>AUC(C2)>AUC(C1).

CS 40003: Data Analytics 60

A Quantitative Measure of a Classifier
The
   concept of ROC plot can be extended to compare quantitatively using
Euclidean distance measure.

 See the following figure for an explanation.

 Here, C(fpr, tpr) is a classifier and denotes the Euclidean distance between
the best classifier (0, 1) and C. That is,


CS 40003: Data Analytics 61

A Quantitative Measure of a Classifier
 The smallest possible value of is 0
 
 The largest possible values of i(when (fpr = 1 and tpr = 0).

 We could hypothesise that the smaller the value of , the better the classifier.

 is a useful measure, but does not take into account the relative importance of true and false positive
rates.

 We can specify the relative importance of making TPR as close to 1 and FPR as close 0 by a weight w
between 0 to 1.

 We can define weighted (denoted by ) as

Note
 If w = 0, it reduces to = fpr, i.e., FP Rate.

 If w = 1, it reduces to = 1 – tpr, i.e., we are only interested to maximizing TP Rate.

CS 40003: Data Analytics 62

 Reference

The detail material related to this lecture can be found in

Data Mining: Concepts and Techniques, (3rd Edn.), Jiawei Han, Micheline Kamber, Morgan
Kaufmann, 2015.

Introduction to Data Mining, Pang-Ning Tan, Michael Steinbach, and Vipin Kumar, Addison-
Wesley, 2014

CS 40003: Data Analytics 63

 Any question?

You may post your question(s) at the “Discussion Forum”

maintained in the course Web page!

CS 40003: Data Analytics 64