TU Wien:Business Intelligence VU (Tjoa)/2. Test - Data Mining - old tests

Test 2012

Q1 Single Choice

1. In k-fold cross validation each data sample is used k-1 times for testing

2. Ordinal data can only be tested for equality/inequality.

3. In decision tree each node denotes a test on an attribute value and each branch represents an outcome of the test.

4. Result of k-means depends on initial guesses of the seeds.

5. Partitional methods allow clusters to be found at different level of granularity.

6. Support vector machines can only be used for linearly separated data.

7. Support Vector Machines with a linear kernel work particularly well with high-dimensional data.

8. 1-n coding doubles the dimension of the feature space.

9. Binning turns ratio quantity attribute into ordinal attribute.

10. High sparsity in feature spaces denotes the concept of insufficient number of training/test data

11. Decision tree pruning is necessary to keep computational load in manageable order of magnitude.

12. K-means is based on supervised leaning paradigm.

13. SOM provides topology-preserving mapping

14. K-means k___to guarantee dec arbitrary dimension of flat space

15. K-anonymity achieved by using high value for k in k-nearest neighbors classification.

Q2

Visualize the different concepts of computing distances between clusters in the figures (like in the slides, MIN, MAX, ...)

Q3

List 3 different types of scaling and describe their characteristics, specifically when they lead to drastically different results and in which case you would apply which type of scaling.

Q4

The following table presents the data of weather conditions, and whether one should play golf given these conditions. Based on this training data, classify the given test data using the Naïve Bayes Classifier.

Test 2013

Q1: Single Choice

1. A lazy learner creates a complex model of the decision space to speed up classification of new data samples.

2. 1-n coding doubles the dimensionality of the feature space.

3. In a decision tree, each node denote a test on an attribute value and each branch represents an outcome of the test.

4. K-anonymity requires at k attributes to uniquely identify a specific data item.

5. High sparsity in feature spaces denotes the concept of insufficient number of training/test data

6. Support Vector Machines can only be used for linearly seperable data

7. Support Vector Machines with a linear kernel work particularly well with high-dimensional data

8. Ordinal data can only be tested for equality/inequality

9. Binning turns ratio quantity attributes into ordinal attributes

10. Partitional methods for clustering allow clusters to be found at different levels of granularity

11. K-means is based on a supervised learning paradigm

12. Decision tree pruning is necessary to keep the computational load in a manageable order of magnitude

13. The Self-Organizing-Map (SOM) is based on a supervised learning paradigm

14. In k-fold cross-validation, each data sample is used k-1 times for testing

15. The result of the k-means method depends on the initial guesses of the seeds

Q2

Visualize the different concepts of computing distances between clusters in the figures (like in the slides, MIN, MAX, ...)

Q3

List 3 different types of scaling and describe their characteristics, specifically when they lead to drastically different results and in which case you would apply which type of scaling.

Q4

The following table presents the data of weather conditions, and whether one should play golf given these conditions. Based on this training data, classify the given test data using the Naïve Bayes Classifier.

Test 2014

Q1: Single Choice

1. Binning describes the process of grouping class labels in multiple layers for hierarchical classification.

2. With 1-n coding of nominal attributes min/max scaling is required so that proper Euclidean distances can be calculated.

3. To minimize impact, missing values can be replaced by ‘0’ when training a Naïve Bayes classifier.

4. The test set is used for selecting the optimal parameter in classifier training

5. Support Vector machines cannot be overtrained, i.e. they do not need a parameter to ensure generalization.

6. In leave-one-out validation, only 1 data sample is used for training case fore each iteration of classifier training.

7. In k-means k has to be an even number to guarantee a decision can be made under arbitrary dimensionality of the feature space.

8. In a Decision Tree, the entropy of the data represented by each branch decreases as we move towards the branches of the tree.

9. Regression is a machine learning task where the target attribute is numeric.

10. K-nearest neighbors is based on a supervised learning paradigm.

11. In a decision tree, each node denotes a test on an attribute value and each branch represents an outcome of the test.

12. Ordinal data shows distances to be computed.

13. Association rules are based on an unsupervised learning paradigm.

Q2

Suppose that we want to determine if someone has an illness ‘C’ based on his symptoms and the given training data in the table below. Each row of the table represents one person and the features f1, … , f4 indicate if the person has a particular symptom or not. The information for the illness is given in column C. You are given these symptoms of a new person <f1,f2,f3,f4>: <0,0,1,0>. Determine if this person has the illness ‘C’ based on k-nearest neighbor, with k=3 distance metric: Hamming or Euclidian distance, or any well-known distance metric of your choice if specified) (Table is missing, but the only values are 0 and 1 for f1 to f4)

Q3

Explain the concept of Data Management Plans and what they are needed for. List the elements it consists of and explain the motivation of moving toward management plans.

Q4

What is the difference between micro-averaging and macro-averaging in classifier evaluation and under which conditions it is important to choose one over the other.

Q5

The following artificial data set represent the characteristics for particular meals, and wheter they are healthy or not. Based on this training data, classify the given test data using Naïve Bayes classifier. The rationale for your decision, i.e. the calculation leading to the result must be provided.

Test 2018

Q1: Single Choice

1. K-nearest is based on supervised leaning paradigm

2. The entropy is at its maximum when the cardinalities of classes are equal

3. Decision tree pruning helps against overfitting

4. With 1-n coding of nominal attributes min/max scaling is required so that proper Euclidean distances can be calculated

5. Support Vector machines cannot be overtrained, i.e. they do not need a parameter to ensure generalization

6. Regression is a machine learning task where the target attribute is numeric

7. Ordinal data shows distances to be computed

8. A decision tree is sensitive against new points (stability)

9. CRISP-DM starts with pre-processing and finishes with evaluation

Q2

What is ACM Statement on Algorithmic Transparency and Accountability? Describe it and why is it important for Data Mining? Which parts belong to it?

Q3

What is binning? What is 1-n coding? Describe and give example. In which case should it be used?

Q4

Naive Bayes: Calculation

Test 2019

2018W: Test B - 07.01.2019

Q1: Multiple Choice

5 questions à 4 answers "All-or-nothing" => 25 points

• About Spark, MapReduce, Mlib...

Q2: Single Choice

15 True/False questions, 2 points for right answer, -1 for wrong answer, 0 for no answer given => 30 points

• Binning turns nominal attributes into ordinal attributes
• Ordinal data allows distances to be computed
• Ordinal data does not allow distances to be computed
• Remaining were also somehow similar to questions in previous exams..

Q3

Open question => 20 points

• Give at least 4 out of 7 ACM Statement on Algorithmic Transparency and Accountability. Describe then and why are they important for Data Mining?
• See for example: https://www.acm.org/binaries/content/assets/public-policy/2017_usacm_statement_algorithms.pdf

Q4

Open question => 25 points

• Naive Bayes: Calculation
• Example give: Predict if someone will be playing some sport under the given features.
• Features: Outlook, Temperature, Humidity, Wind
• Comparable to this example: https://www.youtube.com/watch?v=CPqOCI0ahss

2018W: Test A - 21.03.2019

Q1

5 All or nothing questions about Big Data Analysis

Q2

True/False questions about ML/DM, 2 points for right answer, -1 for wrong answer, 0 for no answer given

Q3

Given a matrix with distances between points do Single Linkage and Complete Linkage Clustering (You had to do the first 3 steps for each)

Q4

knn: Given test samples of persons with various attributes and a class label predict the class label of a new sample.