DMTM Lecture 06 Classification evaluation

Prof. Pier Luca Lanzi
Evaluation of Classification Models
Data Mining andText Mining (UIC 583 @ Politecnico di Milano)

Evaluation & Credidibility Issues
• What measure should we use? (accuracy might not be enough!)
• How reliable are the predicted results?
• How much should we believe in what was learned?
§Error on the training data is not a good indicator of
performance on future data
§The classifier was computed from the very same training data,
any estimate based on that data will be optimistic.
§In addition, new data will probably not be exactly
the same as the training data!
2

How to evaluate the performance of a model?
How to obtain reliable estimates?
How to compare the relative
performance among competing models?
Given two equally performing models,
which one should we prefer?
3

Model Evaluation
• Metrics for Performance Evaluation
§How to evaluate the performance of a model?
• Methods for Performance Evaluation
§How to obtain reliable estimates?
• Methods for Model Comparison
§How to compare the performance of competing models?
• Model Selection
§Which model should we prefer?
4

How to evaluate the performance of a model?
(Metrics for Performance Evaluation)
5

Metrics for Performance Evaluation
• Focus on the predictive capability of a model
• Confusion Matrix:
6
PREDICTED CLASS
TRUE
CLASS
Yes No
Yes a (TP) b (FN)
No c (FP) d (TN)
a:TP (true positive)
b: FN (false negative)
c: FP (false positive)
d:TN (true negative)

Metrics for Performance Evaluation
• Most widely-used metric:
7
PREDICTED CLASS
ACTUAL
CLASS
Yes No
Yes a (TP) b (FN)
No c (FP) d (TN)

Sometimes Accuracy is not Enough
• Consider a 2-class problem
§Number of Class 0 examples = 9990
§Number of Class 1 examples = 10
• If model predicts everything to be class 0,
accuracy is 9990/10000 = 99.9 %
• Accuracy is misleading because model does not
detect any class 1 example
8

Cost Matrix 9
C(x): Cost of misclassifying examples of type x
PREDICTED CLASS
ACTUAL
CLASS
Yes No
Yes C(TP) C(FN)
No C(FP) C(TN)

Computing Cost of Classification 10
Cost
Matrix
PREDICTED CLASS
ACTUAL
CLASS
C(.) + -
+ -1 100
- 1 0
Model M1 PREDICTED CLASS
ACTUAL
CLASS
+ -
+ 150 40
- 60 250
Model M2 PREDICTED CLASS
ACTUAL
CLASS
+ -
+ 250 45
- 5 200
Accuracy = 80%
Cost = 3910
Accuracy = 90%
Cost = 4255

Cost-Based Model Evaluation
• Predictive accuracy in this case is defined as
• True Positives, true negatives
§ Safe and risky loans predicted as safe and risky respectively
• False Positive Errors
§ We accept a risky loan which we predicted was safe
§ We are likely not to get the money back
(let’s say on average 30000 euros)
• False Negative Errors
§ We don’t give a safe loan since we predicted it was risky
§ We will loose the interest money
(let’s say on average 10000 euros
11

Cost-Based Model Evaluation
• Two models, similar performance but different types of errors
• First Model
§1576 false positives and 1224 false negatives
§Total cost is 59525443
• Second Model
§Total cost is 53147717 (more than 6 millions saved)
• What if we can change the way our model makes mistakes?
§Total cost becomes 46852283 (more than 12 millions saved)
12

Costs can be used to evaluate
an existing classification models
Or can be used by the algorithm
to guide the search for the model
Some algorithms can use the cost matrix to
build the model (e.g., decision trees)

Precision & Recall
• Alternatives to accuracy, introduced in the area of information
retrieval and search engine
• Precision
§Focuses on the percentage of examples that have been
classified positive examples and are actually positive
§In the information retrieval context represents the percentage
of actually good documents that have been shown as a result.
• Recall
§Focuses on the percentage of positively classified examples
with respect to the number of existing good documents
§In the information retrieval context, recall represents the
percentage of good documents shown with respect to the
existing ones.
18

Cost-Sensitive Measures
• Precision is biased towards C(TP) & C(FP),
• Recall is biased towards C(TP) & C(FN)
• F1-measure is biased towards all except C(TN),
it is high when both precision and recall are reasonably high
19
The higher the precision, the
lower the FPs
The higher the precision, the
lower the FNs
The higher the F1,
the lower the FPs & FNs

How to obtain reliable estimates?
(Methods for Performance Evaluation)
20

Classification Step 1: Split the Data
Between Train and Test Sets
21
Training set
Testing set
data

Classification Step 2: Compute a
Model using the Data in the Training
Set
22
Training set
Testing set
data
Model Builder
Y N

Classification Step 3: Evaluate the
Model using the Test Set
23
Training set
Testing set
data
Model Builder Predictions
Y N
Evaluate
+
-
+
-

Note on Parameter Tuning
• It is important that the test data is not used
in any way to create the classifier
• Some learning schemes operate in two stages:
§Stage 1: builds the basic structure
§Stage 2: optimizes parameter settings
• The test data can’t be used for parameter tuning!
• Proper procedure uses three sets: training data, validation data,
and test data. Validation data is used to optimize parameters
24

Making the most of the data
• Once evaluation is complete, all the data can be used to build the
final classifier
• Generally, the larger the training data the better the classifier
(but returns diminish)
• The larger the test data the more accurate the error estimate
25

The Full Process: Training, Validation,
and Testing
26
Training set
Validation set
data
Model Builder Predictions
Y N
Evaluate
+
-
+
-
Model
Builder
Testing set
Y N
+
-
+
-
Final
Evaluation

Methods for Performance Evaluation
• How to obtain a reliable estimate of performance?
• Performance of a model may depend on
other factors besides the learning algorithm:
§Class distribution
§Cost of misclassification
§Size of training and test sets
27

Methods of Estimation
• Holdout
§Reserve ½ for training and ½ for testing
§Reserve 2/3 for training and 1/3 for testing
• Random subsampling
§Repeated holdout
• Cross validation
§Partition data into k disjoint subsets
§k-fold: train on k-1 partitions, test on the remaining one
§Leave-one-out: k=n
• Stratified sampling
§oversampling vs undersampling
• Bootstrap
§Sampling with replacement
28

Holdout Evaluation & “Small”
Datasets
• Reserves a certain amount for testing and uses the remainder for
training, typically,
§Reserve ½ for training and ½ for testing
§Reserve 2/3 for training and 1/3 for testing
• For small or “unbalanced” datasets, samples might not be
representative
• For instance, it might generate training or testing datasets with few or
none instances of some classes
• Stratified sampling
§Makes sure that each class is represented with approximately equal
proportions in both subsets
29

Repeated Holdout
• Holdout estimate can be made more reliable by repeating the
process with different subsamples
• In each iteration, a certain proportion is randomly selected for
training (possibly with stratification)
• The error rates on the different iterations are averaged to yield
an overall error rate
• Still not optimum since the different test sets overlap
30

Cross-Validation
• First step
§Data is split into k subsets of equal size
• Second step
§Each subset in turn is used for testing and the remainder for training
• This is called k-fold cross-validation and avoids overlapping test sets
• Often the subsets are stratified before cross-validation is performed
• The error estimates are averaged to yield an overall error estimate
31

Ten-fold Crossvalidation 32
1 2 3 4 5 6 7 8 9 10
1 2 3 4 5 6 7 8 9 10
test train
1 2 3 4 5 6 7 8 9 10
testtrain
2 3 4 5 6 7 8 9 10
test train
1
train
… … …
p1
p2
p10
The final performance is computed as the average pi

Cross-Validation
• Standard method for evaluation stratified ten-fold cross-validation
• Why ten? Extensive experiments have shown that this is the best
choice to get an accurate estimate
• Stratification reduces the estimate’s variance
• Even better: repeated stratified cross-validation
• E.g. ten-fold cross-validation is repeated ten times and results are
averaged (reduces the variance)
• Other approaches appear to be robust, e.g., 5x2 crossvalidation
33

Leave-One-Out Cross-Validation
• It is a particular form of cross-validation
§Set number of folds to number of training instances
§I.e., for n training instances, build classifier n times
• Makes best use of the data
• Involves no random subsampling
• Computationally expensive
34

Leave-One-Out and Stratification
• Disadvantage of Leave-One-Out is that:
stratification is not possible
• It guarantees a non-stratified sample because
there is only one instance in the test set!
• Extreme example: random dataset split equally into two classes
§Best inducer predicts majority class
§50% accuracy on fresh data
§Leave-One-Out-CV estimate is 100% error!
35

Bootstraping
• Cross-validation uses sampling without replacement
§The same instance, once selected, can not be selected again
for a particular training/test set
• Bootstrap uses sampling with replacement
§Sample a dataset of n instances n times with
replacement to form a new dataset of n instances
§Use this data as the training set
§Use the instances from the original dataset that don’t occur in
the new training set for testing
36

The 0.632 Bootstrap
• An instance has a probability of 1–1/n of not being picked
• Thus its probability of ending up in the test data is:
• This means the training data will contain approximately 63.2% of
the instances
37

Estimating Error using Bootstrap
• The error estimate on the test data will be very pessimistic, since
training was on just around 63% of the instances
• Therefore, we compute the overall error ε by combining error
on the train and test as,
• The training error gets less weight than the error on the test data
• Repeat process several times with different replacement samples;
average the results
38

How to compare the relative
performance among competing models?
(Model Comparison)
39

How to Compare the Performance of
Two Models?
• Two models,
§Model M1: accuracy = 85%, tested on 30 instances
§Model M2: accuracy = 75%, tested on 5000 instances
• Can we say M1 is better than M2?
• How much confidence can we place on accuracy of M1 and M2?
• Can the difference in performance measure be explained as a
result of random fluctuations in the test set?
40

Comparing Data Mining Schemes
• Which of two learning schemes performs better?
• The answer is obviously domain-dependent
• The typical approach is to compare 10-fold cross-validation (CV)
estimates using the Student’s t-test
• Student’s t-test tells whether the means of two samples are
significantly different
• Use a paired t-test when the individual samples are paired, that is,
when the same CV is applied twice. Use unpaired t-test
otherwise
41

Confidence Interval for Accuracy
• Prediction can be regarded as a Bernoulli trial with two possible
outcomes, correct or wrong
• A collection of Bernoulli trials has a Binomial distribution
• Given,
§The number of correct test predictions S
§The number of test instances N,
§The accuracy f= S/N
• Can we predict the true accuracy of the model from accuracy f?
42

Mean and Variance
• Assume that the accuracy f comes from a Bernoulli distribution
with mean p and variance p(1-p)
• When N is large enough, f follows a normal distribution with
mean p and variance p(1-p)/N
• c% confidence level interval [-z ≤ X ≤z] for a random variable of
mean 0 is given by
• With a symmetric distribution, we have
43

• For instance, if X has mean 0 and standard deviation 1,
the confidence interval for c% is defined by the z values,
• Thus, for the confidence interval for 90%
we have,
• To compute the confidence interval for the
expected accuracy f we need to,
§ Normalize it to a distribution with mean 0 and standard deviation1
§ Compute the confidence interval
§ Remap it to the actual distribution
44
Pr[X ≥z] z
0.1% 3.09
0.5% 2.58
1% 2.33
5% 1.65
10% 1.28
20% 0.84
25% 0.69
40% 0.25

• First, transform f using,
• Compute the confidence interval,
• Solve it for p,
45

Examples
• When f = 75%, N=1000, c=80% (z=1.28), the confidence
interval for the actual performance is [0.732, 0.767]
• When f = 75%, N=100, c=80% (z=1.28), the confidence interval
for the actual performance is [0.691, 0.801]
• Note that the assumption about f being normally distributed only
applies for large values of N (at least 30)
• When f = 75%, N=10, c=80% (z=1.28), the confidence interval
for the actual performance is [0.549, 0.881]
46

William Gosset (1876-1937)
• Obtained a post as a chemist in the Guinness
brewery in Dublin in 1899.
• Invented the t-test to handle small samples for
quality control in brewing
• Wrote under the name "Student”
• Unpaired and paired two-sample t-test
• Unequal/equal sample size
• Unequal/equal sample variance
47

Comparing the Performance of
Two Models (Paired)
• Fix a significance level
• If a difference is significant at the c% level,
there is a (100-c)% chance that the true means differ
• Divide the significance level by two because the test is two-tailed
• Look up the value for z that corresponds to c/2
• Compute the value of t based on the observed performance
estimates for the schemes being compared
• If t is less than –z or greater than z then the difference is
statistically significant
48

“paired” when the estimates
are from the same datasets
“unpaired” when the estimates
are from different datasets
49

How do we know that the difference in
performance is not just due to change?
We computes the odds of it!
Apply the t-test and compute the p-value
The p-value represents the probability
that the reported difference is due by chance

In repeated experiments at this sample size,
how often would you see such a difference
assuming that there is no difference between
the two groups?

run the python notebook dedicated
to classifier comparison

Multiple Testing
• Say that you perform a statistical test with a 0.05 threshold, but
you repeat the test on twenty different observations.
• For example, you want to compare the performance of several
classification algorithms
• Assume that all of the observations are explainable by the null
hypothesis
• What is the chance that at least one of the observations will
receive a p-value less than 0.05?
46

Multiple Testing - Example
• Say that you perform a statistical test with a 0.05 threshold, but
you repeat the test on 20 different observations. Assuming that
all of the observations are explainable by the null hypothesis,
what is the chance that at least one of the observations will
receive a p-value less than 0.05?
• P(making a mistake) = 0.05
• P(not making a mistake) = 0.95
• P(not making any mistake) = 0.9520 = 0.358
• P(making at least one mistake) = 1 - 0.358 = 0.642
• There is a 64.2% chance of making at least one mistake.
47

Bonferroni Correction
• Assume that individual tests are independent.
• Divide the desired p-value threshold by the number of tests
performed.
• Example
§We now have, the threshold set to 0.05/20 = 0.0025.
§P(making a mistake) = 0.0025
§P(not making a mistake) = 0.9975
§P(not making any mistake) = 0.997520 = 0.9512
§P(making at least one mistake) = 1 - 0.9512 = 0.0488
48

Non Parametric Tests
• Do not make any assumption about the distribution of the
variable in the population
• Mann-Whitney U Test
§Nonparametric equivalent of the independent t-test
• Wilcoxon matched-pairs signed rank test
§Used to compare two related groups
61

Mann Whitney U Test
• To compute the Mann
Whitney U:
§Rank the scores in both
groups (together) from
highest to lowest.
§Sum the ranks of the
scores for each group.
§The sum of ranks for each
group are used to make
the statistical comparison.
Income Rank No Income Rank
25 12 27 10
32 5 19 17
36 3 16 20
40 1 33 4
22 14 30 7
37 2 17 19
20 16 21 15
18 18 23 13
31 6 26 11
29 8 28 9
85 125

Wilcoxon Matched-pairs
Signed Rank test
• Step 1
§ Rank the data of both the groups in ascending order
§ If any values are equal average their ranks
• Step 2
§ Add up the ranks in group with smaller sample size
§ If the two groups are of the same size either one may be picked
§ T= sum of ranks in group with smaller sample size
• Step 3
§ Compare this sum with the critical ranges given in table
§ Look up the rows corresponding to the sample sizes of the two groups
§ A range will be shown for the 5% significance level
63

* 17, 18 & 19 are tied hence the ranks are averaged
Non-smokers (n=15) Heavy smokers (n=14)
Birth wt (Kg) Rank Birth wt (Kg) Rank
3.99 27 3.18 7
3.79 24 2.84 5
3.60* 18 2.90 6
3.73 22 3.27 11
3.21 8 3.85 26
3.60* 18 3.52 14
4.08 28 3.23 9
3.61 20 2.76 4
3.83 25 3.60* 18
3.31 12 3.75 23
4.13 29 3.59 16
3.26 10 3.63 21
3.54 15 2.38 2
3.51 13 2.34 1
2.71 3
Sum=272 Sum=163

Nonparametric tests Parametric tests
Nominal data Ordinal data
Ordinal, interval,
ratio data
One group Chi square
goodness of fit
Wilcoxon signed rank test One group t-test
Two unrelated
groups Chi square
Wilcoxon rank sum test,
Mann-Whitney test
Student’s t-test
Two related
groups
McNemar’s test Wilcoxon signed rank test Paired Student’s t-test
K-unrelated
groups Chi square test
Kruskal -Wallis one-way
analysis of variance
ANOVA
K-related
groups
Friedman matched samples
ANOVA with repeated
measurements

T-Test Using R
> j48 <- c(96,93.33,92.00,94.67,96.00,94.00,94.67,94.0,94.67,98)
> zeror <- c(33.33, 33.33, 33.33, 33.33, 33.33, 33.33, 33.33, 33.33, 33.33, 33.33)
> t.test(j48,zeror)
Welch Two Sample t-test
data: j48 and zeror
t = 117.9102, df = 9, p-value = 1.153e-15
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
60.22594 62.58206
sample estimates:
mean of x mean of y
94.734 33.330
68

Probabilistic Classifiers
70

Probabilistic Classifiers &
Classification Thresholds
• Up to now we used logistic regression to predict classifier labels,
however, logistic regression returns a probability
• Given an example xi its predicted class is the label with the largest
probability so it is equivalent to using a threshold of 0.5 to decide
which class to assign to an example
• However, we can use a different threshold and for instance label
as positive only examples we return a 1 only when P(+1|x)>0.75
• This would label as positive only cases for which we are more
confident that should be labeled as positive.
71

How the Classification Threshold
Influence Precision and Recall?
• Suppose we use a near one threshold to classify positive examples
• Then, we will classify as positives only examples for which we are very
confident (this is a pessimistic classifier)
• Precision will be high
§In fact, we are not likely produce false positives
• Recall will be low
§In fact, we are likely to produce more false negatives
72

How the Classification Threshold
Influence Precision and Recall?
• Suppose we use a near zero threshold to classify positive examples
• Then, we will classify everything as positives
(this is an optimistic classifier)
• Precision will be low as we are going to generate the maximum
number of false positives (everything is positive!)
• Recall will be high since by classifying everything as positive we are
going to generate the maximum number of false negatives
73

We can use the threshold to
optimize our precision and recall
a higher the threshold, increases
precision and lower recall
a lower threshold, decreases precision
and increase recall

Run the notebook dedicated
to the precision-recall tradeoff

Precision-Recall Tradeoff Example
• In the notebook, a simple logistic regression model applied to the loans data
returns the confusion matrix below, corresponding to a precision of 0.82 and
recall of 0.99
• By increasing the classification threshold for positive (+1) examples to 0.75 we
obtained a new confusion matrix with precision of 0.86 and a recall of 0.99
• Overall, we reduced the number of false positives (we did not accept risky
loans and we were better at identifying risky loans)
76
Classified -1 Classified +1
Labeled -1 1212 21910
Labeled +1 1057 98283
Classified -1 Classified +1
Labeled -1 10632 12490
Labeled +1 20106 79234

we can explore this tradeoffs
using precision-recall curves

Precision-Recall Curves
• Plot precision as a function of
recall for varying threshold values
• The best classifier would be the
one that has always a precision
equal to one (but never happens)
• More in general classifiers will show of different shapes
• How to decide among more classifiers?
§Use the area under the curve (the nearer to one, the better)
§Use F1 measure
78

Precision-recall curve for the loan dataset (run the python notebook for details)

Receiver Operating
Characteristic (ROC) Curves

ROC Curve
(Receiver Operating Characteristic)
• Developed in 1950s for signal detection theory to analyze signals
• Plot the True Positive Rate (TPR=TP/(TP+FN))
against the False Positive Rate (FPR=FP/(TN+FP))
• Performance of each classifier represented
as a point on the ROC curve
• Changing the classification threshold, sample distribution or cost
matrix changes the location of the point
81

ROC Curve
• (FPR, TPR)
§(0,0): declare everything
to be negative class
§(1,1): declare everything
to be positive class
§(0,1): ideal
• Diagonal line:
§Random guessing
§Below diagonal line,
prediction is opposite
of the true class
82

ROC Curve 83

Using ROC for Model Comparison
• No model consistently outperform the other
• M1 is better for small FPR
• M2 is better for large FPR
• Area Under the ROC curve
• Ideal, area = 1
• Random guess, area = 0.5
84

Lift Charts
• Lift is a measure of the effectiveness of a predictive model
calculated as the ratio between the results obtained with and
without the predictive model
• Cumulative gains and lift charts are visual aids for measuring
model performance
• Both charts consist of a lift curve and a baseline
• The greater the area between the lift curve and the baseline, the
better the model
89

Example: Direct Marketing
• Mass mailout of promotional offers (1000000)
• The proportion who normally respond is 0.1%, that is 1000
responses
• A data mining tool can identify a subset of a 100000 for which
the response rate is 0.4%, that is 400 responses
• In marketing terminology, the increase of response rate is known
as the lift factor yielded by the model.
• The same data mining tool, may be able to identify 400000
households for which the response rate is 0.2%, that is 800
respondents corresponding to a lift factor of 2.
• The overall goal is to find subsets of test instances that have a
high proportion of true positive
90

An Example from Direct Marketing 91

Which model should we prefer?
(Model selection)
92

Model Selection Criteria
• Model selection criteria attempt to find a good compromise
between:
§The complexity of a model
§Its prediction accuracy on the training data
• Reasoning: a good model is a simple model that achieves high
accuracy on the given data
• Also known as Occam’s Razor :
the best theory is the smallest one
that describes all the facts
93
William of Ockham, born in the village of Ockham in Surrey (England)
about 1285, was the most influential philosopher of the 14th century
and a controversial theologian.

The MDL Principle
• MDL stands for minimum description length
• The description length is defined as:
space required to describe a theory
+
space required to describe the theory’s mistakes
• In our case the theory is the classifier and the mistakes are the
errors on the training data
• We seek a classifier with Minimum Description Language
94

MDL and compression
• MDL principle relates to data compression
• The best theory is the one that compresses the data the most
• I.e. to compress a dataset we generate a model and
then store the model and its mistakes
• We need to compute
(a) size of the model, and
(b) space needed to encode the errors
• (b) easy: use the informational loss function
• (a) need a method to encode the model
95

Which Classification Algorithm?
• Among the several algorithms, which one is the “best”?
§Some algorithms have a lower computational complexity
§Different algorithms provide different representations
§Some algorithms allow the specification of prior knowledge
• If we are interested in the generalization performance,
are there any reasons to prefer one classifier over another?
• Can we expect any classification method to be superior or
inferior overall?
• According to the No Free Lunch Theorem,
the answer to all these questions is known
96

No Free Lunch Theorem
• If the goal is to obtain good generalization performance, there are no
context-independent or usage-independent reasons to favor one
classification method over another
• If one algorithm seems to outperform another in a certain situation, it is
a consequence of its fit to the particular problem, not the general
superiority of the algorithm
• When confronting a new problem, this theorem suggests that we
should focus on the aspects that matter most
§Prior information
§Data distribution
§Amount of training data
§Cost or reward
• The theorem also justifies skepticism regarding studies that
“demonstrate” the overall superiority of a certain algorithm
97

No Free Lunch Theorem
• "[A]ll algorithms that search for an extremum of a cost [objective]
function perform exactly the same, when averaged over all possible
cost functions.“ [1]
• “[T]he average performance of any pair of algorithms across all possible
problems is identical.” [2]
• Wolpert, D.H., Macready, W.G. (1995), No Free Lunch Theorems for
Search, Technical Report SFI-TR-95-02-010 (Santa Fe Institute).
• Wolpert, D.H., Macready, W.G. (1997), No Free Lunch Theorems for
Optimization, IEEE Transactions on Evolutionary Computation 1, 67.
98

DMTM Lecture 06 Classification evaluation

More Related Content

What's hot (20)

Similar to DMTM Lecture 06 Classification evaluation (20)

More from Pier Luca Lanzi (20)

Recently uploaded (20)

DMTM Lecture 06 Classification evaluation