Part D

Predictive Analytics
1 Clustering
D.1.1 Cluster Analysis in python
For the following exercise, load the BigMac2003.csv 1 using pandas. This data set contains for 69 cities world-wide
the average working hours, price level and income level for the year 1991. Your task is to perform a cluster analysis
of the cities based on selected variables. This data set contains the following variables (to list them, you might use,
e.g., attributes(BigMac2003)):
BigMac Minutes of labor to purchase a Big Mac
Bread Minutes of labor to purchase 1 kg of bread
Rice Minutes of labor to purchase 1 kg of rice
FoodIndex Food price index (Zurich=100)
Bus Cost in US dollars for a one-way 10 km ticket
Apt Normal rent (US dollars) of a 3 room apartment
TeachGI Primary teacher’s gross income, 1000s of US dollars
TeachNI Primary teacher’s net income, 1000s of US dollars
TaxRate Tax rate paid by a primary teacher
TeachHours Primary teacher’s hours of work per week

For the class label, generate vector with the cities’ regions (aggregated to 4 classes) as follows:
1 import pandas as pd
2 import numpy as np
3
4 df = pd . read_csv ( " bigMac2003 . csv " )
5 df . rename ( columns ={ df . columns [0]: " City " } , inplace = True )
6
7 # categorical
8 df [ ’ region ’ ]= [ " EU " ," EU " ," AUNZ " ," AS " ," EU " ," EU " ," EU " ," SA " ," EU " ," EU " ," EU " ," EU " ," SA " ," SA " ," NA
",
9 " EU " ," AF " ," EU " ," EU " ," EU " ," EU " ," AS " ," AF " ," AS " ," AF " ," AS " ," EU " ," AS " ," AF " ," SA " ," EU " ," EU " ," EU " ,
" NA " ,
10 " EU " ," EU " ," EU " ," AF " ," AS " ," NA " ," NA " ," EU " ," NA " ," EU " ," AS " ," AF " ," NA " ," EU " ," EU " ," EU " ," EU " ," SA " ,
" EU " ,
11 " SA " ," SA " ," AS " ," AS " ," AS " ," EU " ," EU " ," AUNZ " ," AS " ," EU " ," AF " ," AS " ," NA " ," EU " ," EU " ," EU " ]
12
13 approved = [ " AF " , " AS " , " SA " ]
14 df [ ’ region ’] = np . where ( df [ ’ region ’ ]. isin ( approved ) , df [ ’ region ’] , ’N ’)
15
16 df [ ’ region ’] = df [ ’ region ’ ]. astype ( ’ category ’)
17
18 df . head ()

Now perform the following steps:

1. Select the variables FoodIndex, Bus, Apt, TeachNI and TeachHours and scale them (using the function scale()
to an arithmetic mean of 0 and a standard deviation of 1.
Hint: from sklearn.preprocessing use StandardScaler
Hint: Careful to keep city column as index of the dataframe
1 Provided by the Union Bank of Switzerland.
 2. Compute the distance matrices for distances based on (1 − |correlation|), L1 (manhattan) and L2 (euclidean)
norm:
Hint: use math
Hint: transpose dataframe, calculate correlation using pandas, l1 and l2 with lambda functions
3. Compute an agglomerative hierarchical clustering and analyse the results by studying the resulting dendrograms.
Hint: sklearn provides kmeans, silhouette score and adjusted rand score
Hint: follow this example from scikit-learn
https://scikit-learn.org/stable/auto examples/cluster/plot kmeans silhouette analysis.html
4. Using only the attributes FoodIndex and Apt, compute a k-means based clustering u for k = 3, k = 4, and k = 5,
and analyse the results in a 2-dimensional plot of the clusters and their points. Compute the Silhouette Index
as internal measure for all clustering. For k = 4, compute also the Rand Index as external measure (against the
class label vector region4 ).
Hint: sklearn provides kmeans, silhouette score and adjusted rand score
Hint: follow this example from sklearn https://scikit-learn.org/stable/auto examples/cluster/plot kmeans silhouette analy
5. As above, use the attributes FoodIndex and Apt, but now compute an EM-based clustering for nclass = 3 and
nclass = 5 and analyse the results in a 2-dimensional plot of the clusters and their points. As above, compute
the Silhouette and Rand Indices.
Hint: from sklearn.mixture use GaussianMixture
6. Generate a knee-plot for determining the optimal number of clusters in your kmeans or EM-clusterings based
on the Silhouette Index.

D.1.2 How would you distinguish between flat and hierarchical partitioning?
D.1.3 What is probabilistic clustering?
D.1.4 What is the difference between soft and hard partitioning?
D.1.5 What are the two main approaches of flat partitioning?
D.1.6 What are the two main approaches of hierarchical partitioning?
D.1.7 Discuss the Minkowski-Distances for r = 1 and r = 2, their differences and their
alternative names.
D.1.8 On which of the following datasets would you use (1) K-Means Clustering (or
EM), (2) single-linkage agglomerative hierarchical clustering, (3) density-based
clustering? Provide arguments for your recommendation.
D.1.9 Which clustering(s) tend(s) to form clusters that are chains of points?
( ) Single Linkage Agglomerative Hierarchical Clustering

( ) Complete Linkage Agglomerative Hierarchical Clustering

( ) K-Means Clustering
( ) Expectation Maximisation with Mixture of Gaussians

D.1.10 Which clustering(s) tend(s) to form spherical clusters?

( ) Single Linkage Agglomerative Hierarchical Clustering
( ) Complete Linkage Agglomerative Hierarchical Clustering
( ) K-Means Clustering

( ) Expectation Maximisation with Mixture of Gaussians

2
X2

X2

X2
X1 X1 X1
(a) (b) (c)
X2

X2

X2

X1 X1 X1
(d) (e) (f)

Figure 1: Scatter plots of different data sets.

D.1.11 Which of the following similarity/distance measures is appropriate for compar-

ing instances based on their relative attribute values?
( ) Euclidean Distance
( ) L1 norm or City-Block Metric

( ) Mahanalobis distance
( ) Q-Correlation Coefficient

D.1.12 Which of the following similarity/distance measures is appropriate for compar-

ing instances based on the absolute distance on their attribute values?
( ) Euclidean Distance

( ) L1 norm or City-Block Metric

( ) Mahanalobis distance
( ) Q-Correlation Coefficient

D.1.13 Which of the following statements are correct?

( ) The Silhouette Index is an internal cluster quality measure.
( ) Internal cluster quality measures compare the cluster quality based on ground-truth class labels.
3
 ( ) External cluster quality measures compare the cluster quality based on ground-truth class labels.

( ) Clustering aims to maximise similarity between instances within clusters and dissimilarity between instances in
different clusters.
( ) The Silhouette Index considers the intra-cluster and inter-cluster distances.

D.1.14 You want to compare two clusterings, one with three and one with five clusters.
Which external performance measures can you use?
( ) Purity
( ) Normalised Mutual Information

D.1.15 Based on the elbow method, determine in the plots in Fig. 2 below which
number of clusters k to chose.
RSS

RSS

RSS

1 2 3 4 5 6 7 8 k 1 2 3 4 5 6 7 8 k 1 2 3 4 5 6 7 8 k
(a) (b) (c)

Figure 2: Elbow plots for three different data sets.

4
2 Classification
D.2.1 Which of the following statements are correct?
( ) In classification, the task is to predict a dependent variable (class label) based on a set of explanatory (feature)
variables.

( ) A probabilistic classifier returns also estimates of the posterior class probability Pr (y|x).
( ) A lazy learner does not construct an (abstract) model from the data.
( ) A Bayes classifier requires an estimate of either the joint probability Pr (x, y), or estimates of the class prior
probability Pr (y) together with the class-conditional feature probability Pr (x|y).

( ) The (unconditional) feature probability Pr () can be computed by summing (marginalising) Pr (x, y) over the
different values of x.
( ) The (unconditional) feature probability Pr () can be computed by summing (marginalising) Pr (x, y) over the
different values of y.

( ) In contrast to a multivariate Bayes classifier, a Naive Bayes classifier assumes conditional independence.
( ) A predictive classifier also provides a model of the underlying data distribution, which can be used to generate
data.
( ) For k-fold cross validation, the data set is partitioned into k subsets, and each of them is used once for testing
and the other times for training.
( ) For obtaining a reliable estimate of the performance on new data, we need to evaluate a classifier on its training
data.
( ) False positives are instances that are classified as negative, but actually are positive.

( ) In a region, where we classify all instances as positive, our false negative rate is zero.

D.2.2 Curse of Dimensionality: If you have three feature variables X1 ,X2 , and X3 , which
you divide each into two intervals (low vs. high values), in how many cells does
this split your feature space? Alternatively, how many training instances do you
need to place exactly one in each cell? How does this change with four features
X1 ,X2 ,X3 and X4 ?

5
D.2.3 In Fig. 3, the scatterplots of two data sets (a) and (b) are given. In which of
these two is conditional independence assumption clearly violated?

Negatives Negatives
X2

X2
Positives Positives
P(X2|-)
P(X2|+)

P(X2|+)
P(X2|-)
X1 P(X1|-)
X1
P(X1|-) P(X1|+) P(X1|+)

(a) (b)

Figure 3: Scatterplots of two datasets.

6
D.2.4 Based on the AUC/ROC plotted in Fig. 4, which of the following statement(s)
is/are correct?

True Positive Rate (Sensitivity)

0.5 0.6 0.7 0.8 0.9 1

Classifier A
0.4

Classifier B
0.3

Classifier C
0.2

Classifier D
0.1
0

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

False Positive Rate (1-Specificity)

Figure 4: Area Under the Receiver Operating Characteristics (ROC) Curve (AUC) Plots.

( ) Classifier A performs best.

( ) Classifier A is a perfect classifier.

( ) Classifier D is a perfect classifier.
( ) Classifier A performs like a randomly guessing classifier.

( ) Classifier D performs like a randomly guessing classifier.

( ) Classifier A dominates classifiers B,C,D.
( ) Classifier B dominates classifiers C,D.
( ) Classifier D dominates classifiers A,B,C.

( ) Classifier D is dominated by classifiers A,B, and C.

( ) Neither classifier B dominates classifier C, nor classifier C dominates classifier B.

7
D.2.5 (Univariate) Bayesian Classification
Consider a binary classification problem, with class variable Y ∈ {pos, neg} and feature variable X ∈ (0, 15). The
joint distributions of P (X, Y ) are given in the plot below (P (X, Y = pos) is plotted in green, P (X, Y = neg) in blue).
0.3
P(x,y=+)
P(x,y=-)

0.25
Probability Density Distribution

0.2

0.15

0.1

0.05

0
0 2 4 6 8 10 12 14
x1 x2 x3 x4
Feature X

P (X) P (X, Y = +) P (X, Y = −) Sum

x1 = 4 0.077122 0.158710 0.235830
x2 = 6 0.113980 0.015205 0.129180
x3 = 8 0.035309 0.051915 0.087224
x4 = 12 0.000031 0.051907 0.051939

a) Plot the feature distribution P (X)

b) Can you derive the posterior class distributions P (Y = pos|X) and P (Y = neg|X) from the presented plot?
Sketch them in the plot.

c) Mark the Bayes’ optimal decision boundary/boundaries. How would you classify the following instances: x1 = 4,
x2 = 6, x3 = 8, x4 = 12
d) Indicate the Bayes’ error rate. Assuming a Bayes’ optimal decision, annotate the number of false negatives, and
the number of true negatives in the plot.

8
D.2.6 Multivariate Bayesian Classification
Consider the following, simplified golf-player classification example2 :
Temp Humidity Windy Play
hot high false no
hot high true no
hot high false yes
mild high false yes
cool normal false yes
cool normal true no
cool normal true yes
mild high false no
cool normal false yes
mild normal false yes
mild normal true yes
mild high true yes
hot normal false yes
mild high true no

a) Calculate and interpret the following probabilities:

• Pr(Xtemp = hot)
• Pr(y = yes)
• Pr(Xtemp = hot, y = yes)
• Pr(Xtemp = hot|y = yes)
• Pr(y = yes|Xtemp = hot)

b) Explain the classification of a multivariate Bayes-optimal classifier for an instance with Xtemp = cool and
Xhum = normal and calculate the necessary probabilities.

c) Explain the classification of a Naive Bayes classifier for an instance with Xtemp = cool and Xhum = normal
and calculate the necessary values.

2 Based on a fictitious dataset from [?, chapter 1]

9
 D.2.7 Classification and Regression in python
In python, replicate the Golf-Player data set from the exercise above with the following command:
1
2 temp = [ " hot " ," hot " ," hot " ," mild " ," cool " ," cool " ," cool " ," mild " ," cool " ," mild " ," mild " ," mild " ,"
hot " ," mild " ]
3 hum = [ " h " ," h " ," h " ," h " ," n " ," n " ," n " ," h " ," n " ," n " ," n " ," h " ," n " ," h " ]
4 windy = [ " f " ," t " ," f " ," f " ," f " ," t " ," t " ," f " ," f " ," f " ," t " ," t " ," f " ," t " ]
5 play = [ " n " ," n " ," y " ," y " ," y " ," n " ," y " ," n " ," y " ," y " ," y " ," y " ," y " ," n " ]
6 d = { ’ temp ’: temp , ’ hum ’: hum , ’ windy ’: windy , ’ play ’: play }
7 golf_traine = pd . DataFrame ( data = d )
8 golf_traine . head ()
Next, generate a test set with the following commands:
1
2 temp = [ " cool " ," cool " ," hot " ]
3 hum = [ " n " ," n " ," h " ]
4 windy = [ " t " ," f " ," t " ]
5 play = [ " n " ," y " ," n " ]
6 d = { ’ temp ’: temp , ’ hum ’: hum , ’ windy ’: windy , ’ play ’: play }
7 golf_test = pd . DataFrame ( data = d )
8 golf_test . head ()

a) In python, train a Naive Bayes classifier on the attributes temp and hum, and to predict the class label of the
test instances. Compare the results with your calculation in the exercise above!
Hint: from sklearn.naive bayes use GaussianNB
Hint: from sklearn use preprocessing
b) In python,train a decision tree on the attributes temp and hum, and to predict the class label of the test in-
stances. Compare the results with above!
Hint: from sklearn.tree use DecisionTreeClassifier

c) Visualize the tree and see how it decides how an instance is classified
Hint: plot tree
d) In this exercise you will get familiar with AutoML. What is AutoML? Automated Machine Learning (AutoML)
is tied in with producing Machine Learning solutions for the data scientist without doing unlimited inquiries on
data preparation, model selection, model hyperparameters and model compression parameters.

One popular implementation of AutoML is auto-sklearn:

https://automl.github.io/auto-sklearn/master/index.html

Find the best classifier for the breast cancer dataset from scikit-learn by following this example:
https://automl.github.io/auto-sklearn/master/examples/20 basic/example classification.html#sphx-glr-examples-2
0-basic-example-classification-py

3 Frequent Itemset Mining

D.3.1 Frequent Itemset Mining
Apply the apriori algorithm to the basket data in Table 1, using as threshold for support σ ≥ 38 and for confidence c ≥ 21 .
Based on the algorithm, which recommendations (in descending order of relevance; also state the recommendations’
confidences!) should be made to the following customers:
a) Customer I, who has bought B (single item recommendation suffices).
b) Customer II, who has bought B and C (single item recommendation suffices).

10
IID Item Name TID Items in Baskez
A Grey’s Anatomy 1 A,B,C,D,G,H
B The Big Bang Theory 2 A,B,C,G
C Castle 3 A,D,H
D Downton Abbey 4 B,C
G Game of Thrones 5 B,C,D,G
H How I Met Your Mother 6 B,C,G
7 B,D,G
8 B,G
(a) List of Items (b) Transactions / Baskets

Table 1: Items and Baskets for Frequent Itemset Mining

11
Appendix

12