Which is different from the others:

• Expectation Maximisation
• Apriori
• K-means
• Decision Tree

What is the cross validation:

• A technique to obtain a good estimation of the performance of a classifier when it will be used with data different from the training set
• A technique to obtain a good estimation of the performance of a classifier with the training set
• A technique to improve the quality of a classifier
• A technique to improve the speed of a classifier

What is the single linkage?

• A method to compute the distance between two objects, it can be used in hierarchical clustering
• A method to compute the distance between two sets of items, it can be used in hierarchical clustering
• A method to compute the distance between two classes, it can be used in decision trees
• A method to compute the separation of the objects inside a cluster

Which of the following characteristic of data can reduce the effectiveness of DBSCAN?

• Clusters have concavities
• All the variables are the same range of values
• Presence of outliers
• Presence of clusters with different densities

Which of the following types of data allows the use of the euclidean distance?

• Points in a vector space
• Ordered data
• Transactional data
• Document representations

How does pruning work when generating frequent itemsets?

• If an itemset is not frequent, then none of its supersets can be frequent, therefore the frequencies of the supersets are not evaluated
• If an itemset is frequent, then none of its supersets can be frequent, therefore the frequencies of the supersets are not evaluated
• If an itemset is not frequent, then none of its subsets can be frequent, therefore the frequencies of the subsets are not evaluated
• If an itemset is frequent, then none of its subsets can be frequent, therefore the frequencies of the subsets are not evaluated

In a decision tree, the number of objects in a node…

• … is smaller than the number of objects in its ancestor
• … is not related to the number of objects in its ancestor
• … is smaller than or equal to the number of objects in its ancestor
• … is bigger than the number of objects in its ancestor

What does K-means try to minimise?

• The distortion, that is the sum of the squared distances of each point with respect to the points of the other clusters
• The distortion, that is the sum of the squared distances of each point with respect to its centroid
• The separation, that is the sum of the squared distances of each point with respect to its centroid
• The separation, that is the sum of the squared distances of each cluster centroid with respect to the global centroid of the dataset

Given the following definitions: TP = True Positives; TN = True Negatives; FP = False Positives; FN = False Negatives. Which of the formulas below computes the precision of a binary classifier?

• $TP/(TP+FP)$
• $(TP+TN)/(TP+FP+TN+FN)$
• $TN/(TN+FP)$
• $TP/(TP+FN)$

Given the two binary vectors below, which is their similarity according to the Jaccard Coefficient?

1000101101
1011101010

• 0.1
• 0.2
• 0.5
• 0.375

Consider the transactional dataset below: Which is the support of the rule A,C→B

ID Items
A,B,C
A,B,D
B,D,E
C,D
A,C,D,E

• 100%
• 20%
• 40%
• 50%

Which of the following statements is true?

• The noise can generate outliers
• The noise always generate outliers
• The data which are similar to the majority are never noise
• Outliers can be due to noise

Which of the following clustering methods is not based on distances

• between objects?
• Hierarchical Agglomerative
• DBSCAN
• K-Means
• Expectation Maximization

In a decision tree, an attribute which is used only in nodes near the leaves…

• … is irrelevant with respect to the target
• … guarantees high increment of purity
• … gives little insight with respect to the target
• … has a high correlation with respect to the target

Which of the statements below is true? (Only one)

• Sometimes k-means stops to a configuration which does not give the minimum distortion for the chosen value of the number of clusters
• K-means always stop to a configuration which gives the minimum distortion for the chosen value of the number of clusters
• K-means works well also with datasets having a very large number of attributes
• K-means finds the number of clusters which give the minimum distortion

Which of the statements below is true? (One or more)

• Increasing the radius of the neighourhood can decrease the number of noise points
• DBSCAN always stops to a configuration which gives the optimal number of clusters
• Sometimes DBSCAN stops to a configuration which does not include any cluster
• DBScan can give good performance when clusters have concavities

In data preparation which is the effect of normalisation?

• Map all the numeric attributes to the same range, without altering the distribution, in order to avoid that attributes with large ranges have more influence
• Map all the numeric attributes subtracting the respective mean and dividing by variance
• Map all the numeric attributes to the same range and to have a Gaussian (or normal distribution, in order to avoid that attributes with large ranges have more influence
• Map all the numeric attributes in order to have a Gaussian (or 𝘯𝘰𝘳𝘮𝘢𝘭## 𝘥𝘪𝘴𝘵𝘳𝘪𝘣𝘶𝘵𝘪𝘰𝘯, in order to avoid that attributes with large ranges have more influence

Consider the transactional dataset below: Which is the confidence of the rule A,C→B

ID Items
A,B,C
A,B,D
B,D,E
C,D
A,C,D,E

• 50%
• 40%
• 20%
• 100%

Which of the statements below about Hierarchical Agglomerative Clustering is true?

• Requires the definition of Inertia of clusters
• Requires the definition of distance between sets of objects
• Is very efficient also with large datasets
• Is based on a well founded statistical model

When training a neural network, what is the learning rate?

• The speed of convergence to a stable solution during the learning process
• The ratio between the size of the hidden layer and the input layer of the network
• The slope of the activation function in a specific node
• A multiplying factor of the correction to be applied to the connection weights

What are the hyperparameters of a Neural Network? (Possibly non exhaustive)

• Hidden layers structure, Output layer structure, Activation function, Number of epochs
• Network structure, Learning rate, Backpropagation algorithm, Number of epochs
• Input layers structure, Learning rate, Activation function, Number of epochs
• Hidden layers structure, Learning rate, Activation function, Number of epochs

Which of the following preprocessing activities is useful to build a Naive Bayes classifier if the independence hypothesis is violated:

• Discretisation
• Normalisation
• Standardisation
• Feature selection

Which is different from the others?

• Silhouette Index
• Misclassification Error
• Entropy
• Gini Index

Why do we prune a decision tree?

• To eliminate parts of the tree where the decisions could be influenced by random effects
• To eliminate rows of the dataset which could be influenced by random effects
• To eliminate parts of the tree where the decision could generate underfitting
• To eliminate attributes which could be influenced by random effects

In a dataset with D attributes, how many subsets of attributes should be considered for feature selection according tp an exhaustive search?

• $o(2^d)$
• $o(D!)$
• $o(D)$
• $o(D^2)$

In which mining activity the Information Gain can be useful?

• Discovery of association rules
• Classification
• Clustering
• Discretization

Given the two binary vectors below, which is their similarity according to the Simple Matching Coefficient?

1000101101
1011101010

• 0.1
• 0.3
• 0.5
• 0.2

Which is the main reason for the standardization of numeric attributes?

• Change the distribution of the numeric attributes, in order to obtain gaussian distributions
• Map all the numeric attributes to a new range such that the mean is zero and the variance is one
• Map all the nominal attributes to the same range, in order to prevent the values with higher frequency from having prevailing influence
• Remove non-standard values

Which of the following measure can be used as an alternative to the Information Gain?

• Sihlouette Index
• Gini Index
• Rand Index
• Jaccard Index

Which of the following is not an objective of feature selection

• Select the features with higher range, which have more influence on the computation
• Reduce time and memory complexity of the mining algorithms
• Reduce the effect of noise
• Avoid the curse of dimensionality

After fitting DBSCAN with the default parameter values the results are: 0 clusters, 100% of noise points. Which will be your next trial?

• Reduce the minimum number of objects in the neighborhood
• Reduce the minimum number of objects in the neighborhood and the radius of the neighborhood
• Increase the radius of the neighborhood
• Decrease the radius of the neighborhood

What is the meaning of the statement “the support is anti-monotone”?

• The support of an itemset never exceeds the support if its subsets
• The support of an itemset is always smaller than the support of its supersets
• The support of an itemsets never exceeds the support if its supersets
• The support of an itemset is always smaller than the support of its subsets

In feature selection, what is the Principal Component Analysis?

• A mathematical technique used to transform non numeric attributes into numeric attributes
• A mathematical technique used to transform a set of numeric attributes into a smaller set of numeric attributes which capture most of the variability in data
• A heuristic technique used to find a subset of the attributes which produces the same classifier
• A mathematical technique used to find the principal attributes which determine the classification process

A Decision Tree is…

• A tree-structured plan of tests on single attributes to forecast the cluster
• A tree-structured plan of tests on single attributes to forecast the target
• A tree-structured plan of tests on single attributes to obtain the maximum purity of a node
• A tree-structured plan of tests on multiple attributes to forecast the target

In data preprocessing, which of the following are the objectives of the aggregation of attributes?

• Obtain a more detailed description of data
• Obtain a less detailed scale
• Reduce the variability of data
• Reduce the number of attributes or distinct values

In order to reduce the dimensionality of a dataset, which is the advantage of Multi Dimensional Scaling (MDS), with respect to Principal Component Analysis (PCA)

• MDS requires less computational power
• MDS can be used also with categorical data, provided that the matrix of the distance is available, while PCA is limited to vector spaces
• MDS can be used with categorical data after a transformation in a vector space
• MDS can work on any kind of data, while PCA is limited to categorical data

What is the main purpose of smoothing in Bayesian classification?

• Classifying an object containing attribute values which are missing from some classes in the training set
• Dealing with missing values
• Reduce the variability of the data
• Classifying an object containing attribute values which are missing from some classes in the test set

Consider the transactional dataset below: Which is the confidence of the rule B → E ?

ID Items
A,B,C
A,B,D
B,D,E
C,D
A,C,D,E

• 100%
• 50%
• 33%
• 20%

In a Decision Tree for classification, what is a lead node?

• A node which assigns a class value to the objects passing the tests on the path from the root to the node itself
• A node where all the objects belong to the same class
• A node which allows classification without errors
• A node which assigns a class value only by majority of examples

Match the rule evaluation formulas with their names:

• 1. $sup(A\cup C)-sup(A)sup(C)$
• 2. $\frac{conf(A\Rightarrow C)}{sup(C)}$
• 3. $\frac{sup(A\Rightarrow C)}{sup(A)}$
• 4. $\frac{1-sup(C)}{1-conf(A\Rightarrow C)}$ ​ Options:

• Confidence → 3

• Lift → 2

• Leverage → 1

• Conviction → 4

Which of the statements below best describes the strategy of Apriori in finding the frequent itemsets?

• Evaluation of the support of the itemsets in an order such that uninteresting parts of the search space are considered only at the end of the execution
• Evaluation of the confidence of the itemsets in an order such that uninteresting parts of the search space are pruned as soon as possible
• Evaluation of the support of the itemsets in an order such that uninteresting parts of the search space are pruned as soon as possible
• Evaluation of the support of the itemsets in an order such that the interesting parts of the search space are pruned as soon as possible

What is the Gini index?

• A measure of the entropy of a dataset
• An accuracy measure of a dataset alternative to the Information Gain and to the Misclassification Index
• An impurity measure of a dataset alternative to the Information Gain and to the Misclassification Index
• An impurity measure of a dataset alternative to overfitting and underfitting

What measure is maximised by the Expectation Maximisation algorithm for clustering?

• The likelihood of a class label, given the attributes of the example
• The likelihood of an attribute, given the class label
• The likelihood of an example
• The support of a class

Given the following definitions: TP = True Positives; TN = True Negatives; FP = False Positives; FN = False Negatives. Which of the formulas below computes the accuracy of a binary classifier?

• $\frac{TN}{TN+FP}$
• $\frac{TP+TN}{TP+FP+TN+FN}$
• $\frac{TP}{TP+FN}$
• $\frac{TP}{TP+FP}$

answer: 2

In data preprocessing, which of the following is not an objective of the aggregation of attributes

• Reduce the variability of data
• Obtain a less detailed scale
• Reduce the number of attributes or objects
• Obtain a more detailed description of data

Which of the following is a strength of the clustering algorithm DBSCAN?

• Requires to set the number of clusters as a parameter
• Ability to find cluster with concavities
• Ability to separate outliers from regular data
• Very fast by computation

Which of the following statements regarding the discovery of association rules is true? (One or more)

• The confidence of a rule can be computed starting from the supports of the itemsets
• The support of an itemset is anti-monotonic with respect to the composition of the itemset
• The support of a rule can be computed given the confidence of the rule
• The confidence of an itemset is anti-monotonic with respect to the composition of the itemset

For each type of data choose the best suited distance function

• Vector space with real values → euclidean distance
• Vectors of terms representing documents → cosine distance
• High dimensional spaces → Manhattan distance
• Boolean data → jaccard coefficient

Which of the following is not a strength point of Dbscan with respect to K-means

• The robustness with respect to outliers
• The effectiveness, even in presence of noise
• The effectiveness even if there are clusters with non-convex shape
• The efficiency even in large datasets

Which is the effect of the curse of dimensionality

• When the number of dimensions increases the results tend to be prone to overfitting
• When the number of dimensions increases the euclidean distance becomes less effective to discriminate between points in the space
• When the number of dimensions increases the computing power necessary to compute the distances becomes too high
• When the number of dimensions increases the classifiers cannot be correctly tuned

Which is different from the others?

• Expectation Maximisation
• Decision Tree
• K-means
• Dbscan

In a Neural Network, what is the backpropagation?

• The technique used to adjust the node weights according to the difference between the desired output and the output generated by the network
• The technique used to adjust the weights limiting the probability of overfitting
• The technique used to adjust the output according to the difference between the desired weights and the actual weights
• The technique used to adjust the connection weights according to the difference between the desired output and the output generated by the network

Which is different from the others?

• Dbscan
• SVM
• Neural Network
• Decision Tree

Which of the following is not a property of a metric distance function

• Symmetry
• Positive definiteness
• Triangle inequality
• Boundedness

Which of the statements below is true? (One or more)

• K-mean always stops to a configuration which gives the minimum distortion for the chosen value of the number of clusters
• K-means is very sensitive to the initial assignment of the centers
• K-means is quite efficient even for large datasets
• Sometimes k-means stops to a configuration which does not give the minimum distortion for the chosen value of the number of clusters

Which of the following characteristic of data can reduce the effectiveness of K-Means?

• Presence of values with high frequency
• Presence of outliers
• All the variables have the same distribution of values
• All the variables have the same range of values

Which is the purpose of discretisation?

• Increase the number of distinct values in an attribute, in order to put in evidence possible patterns and regularities
• Reduce the number of distinct values in an attribute, in order to increase the efficiency of the computations
• Reduce the number of distinct values in an attribute, in order to put in evidence possible patterns and regularities
• Reduce the range of values of a numeric attribute, to make all the attributes more comparable

Given the following definitions: TP = True Positives; TN = True Negatives; FP = False Positives; FN = False Negatives. Which of the formulas below computes the recall of a binary classifier?

• $\frac{TP}{TP+FN}$
• $\frac{TN}{TN+FP}$
• $\frac{TP}{TP+FP}$
• $\frac{TP+TN}{TP+FP+TN+FN}$

answer: 1

The Information Gain is used to

• select the attribute which maximises, for a given training set, the ability to predict the class value
• select the attribute which maximises, for a given test set, the ability to predict the class value
• select the attribute which maximises, for a given training set, the ability to predict all the other attribute values
• select the class with maximum probability

Which is the main reason for the MinMax scaling (also known as “rescaling”) of attributes?

• Change the distribution of the numeric attributes, in order to obtain gaussian distributions
• Map all the nominal attributes to the same range in order to prevent the values with higher frequency from having prevailing influence
• Remove abnormal values
• Map all the numeric attributes to the same range, in order to prevent attributes with higher range from having prevalent influence

Which of the following is a base hypothesis for a bayesian classifier?

• The attributes must be statistically independent inside each class
• The attributes must have negative correlation
• The attributes must be statistically independent
• The attributes must have zero correlation

When developing a classifier, which of the following is a symptom of overfitting?

• The error rate in the test set is much smaller than the error rate in the training set
• The precision is much greater than the recall
• The error rate in the test set is more than 30%
• The error rate in the test set is much greater than the error rate in the training set

With reference to the total sum of squared errors and separation of a clustering scheme, which of the statements below is true?

• They are two ways to measure the same thing
• They are strictly correlated, if, changing the clustering scheme, one increases, then the other does the same
• It is possible to optimise them (i.e. minimise SSE and maximise SSB) separately
• They are strictly correlated, if, changing the clustering scheme, one increases, then the other decreases