Corrected Uppsala University Department of Information Technology Data Mining - exam questions These questions have been prepared as material for the Data Mining course at Uppsala University. Please do not share them with people not enrolled in the course. The written exam will contain a selection of these (or similar) questions. 4 Classification Question 1 Consider the following confusion matrix: True 0 True 1 Predicted 0 35 30 Predicted 1 10 30 What is the accuracy of the classifier? A .35 B .78 C .62 D .54 E None of the previous answers Question 2 Consider the following confusion matrix: True 0 True 1 Predicted 0 35 30 Predicted 1 10 25 What is the accuracy of the classifier? A .78 B .35 C .6 D .54 E None of the previous answers Question 3 Consider the following confusion matrix: True 0 True 1 Predicted 0 35 20 Predicted 1 20 25 What is the accuracy of the classifier? A .6 B .78 C .54 D .35 E None of the previous answers Corrected Question 4 Consider the following confusion matrix: True 0 True 1 Predicted 0 35 20 Predicted 1 20 30 What is the accuracy of the classifier? A .78 B .54 C .62 D .35 E None of the previous answers Question 5 Consider the following confusion matrix: True 0 True 1 Predicted 0 35 30 Predicted 1 20 25 where 0 is the positive class. What is the recall of the classifier? A .35 B .54 C .64 D .78 E None of the previous answers Question 6 Consider the following confusion matrix: True 0 True 1 Predicted 0 35 30 Predicted 1 20 25 where 0 is the positive class. What is the recall of the classifier? A .78 B .44 C .35 D .74 E None of the previous answers Question 7 Consider the following confusion matrix: True 0 True 1 Predicted 0 35 20 Predicted 1 20 25 where 0 is the positive class. What is the recall of the classifier? A .52 B .33 C .64 D .78 E None of the previous answers Corrected Question 8 Consider the following confusion matrix: True 0 True 1 Predicted 0 35 30 Predicted 1 10 25 where 0 is the positive class. What is the recall of the classifier? A .6 B .54 C .78 D .35 E None of the previous answers Question 9 Consider the following confusion matrix: True 0 True 1 Predicted 0 35 30 Predicted 1 10 25 where 0 is the positive class. What is the precision of the classifier? A .6 B .54 C .35 D .78 E None of the previous answers Question 10 Which of the following is the definition of precision? A TP T P +F P +T N +F N B TP F P +T N C TP T P +F N D TP F P +F N E TP T P +T N F None of the previous answers Question 11 Predicted 0 Predicted 1 Consider True 0 35 10 the following confusion matrix: True 1 25 30 where 0 is the positive class. What is the precision of the classifier? A .25 B .78 C .35 D .68 E None of the previous answers Corrected Question 12 Which of the following is the definition of precision? A TP T P +F P B TP F P +F N C TP T P +T N D TP T P +F P +T N +F N E TP F P +T N F None of the previous answers Question 13 Predicted 0 Predicted 1 Consider True 0 35 10 the following confusion matrix: True 1 30 30 where 0 is the positive class. What is the precision of the classifier? A .78 B .6 C .54 D .35 E None of the previous answers Question 14 Predicted 0 Predicted 1 Consider True 0 35 10 the following confusion matrix: True 1 25 30 where 0 is the positive class. What is the precision of the classifier? A .58 B .78 C .6 D .35 E None of the previous answers Question 15 How is the entropy of a node t defined, as used in the C4.5 algorithm? (in the following definitions C is the set of all classes and p(c|t) is the frequency of class c at node t) P A c∈C 1 − max(p(c|t)) P B c∈C p(c|t) log p(c|t) P C c∈C log p(c|t) P D c∈C 1 − min(p(c|t)) P E − c∈C p(c|t) log p(c|t) P F c∈C max(p(c|t)) P G c∈C p(c|t) H None of the previous answers Corrected Question 16 How is the entropy of a node t defined, as used in the C4.5 algorithm? (in the following definitions C is the set of all classes and p(c|t) is the frequency of class c at node t) P A c∈C p(c|t) P B c∈C 1 − max(p(c|t)) P C c∈C log p(c|t) P D c∈C max(p(c|t)) P E c∈C p(c|t) log p(c|t) P F c∈C 1 − min(p(c|t)) P G − c∈C min(p(c|t)) H None of the previous answers Question 17 Consider a decision tree where a node has been split into two leaves. The first leaf contains 5 records, 2 of class c0 and 3 of class c1. The second leaf contains 5 records, 4 of class c0 and 1 of class c1. What is the classification error of this split? A .25 B .55 C .15 D .30 E .40 F None of the previous answers Question 18 Consider a decision tree where a node has been split into two leaves. The first leaf contains 5 records, 3 of class c0 and 2 of class c1. The second leaf contains 5 records, 5 of class c0 and 0 of class c1. What is the classification error of this split? A .3 B .2 C .1 D .4 E .5 F None of the previous answers Question 19 Consider a decision tree where a node has been split into two leaves. The first leaf contains 4 records, 2 of class c0 and 2 of class c1. The second leaf contains 4 records, 4 of class c0 and 0 of class c1. What is the classification error of this split? A .25 B .30 C .15 D .40 E .55 F None of the previous answers Corrected Question 20 Which of the following methods partitions the dataset into a training and a test set (that is, each record is used only once and is included either in the training or in the test set)? A k-fold validation B Boosting C Bagging D Leave one out E Bootstrap F None of the previous answers Question 21 Which of the following methods partitions the dataset into a training and a test set (that is, each record is used only once and is included either in the training or in the test set)? A Leave one out B Bootstrap C Cross validation D Holdout E k-fold validation F None of the previous answers Question 22 In boosting, records that are wrongly classified in previous rounds A do not change their probability of being included in the test set. B do not change their probability of being included in the training set. C always have their weights increased. D always have their weights decreased. E None of the previous answers Question 23 In bagging, records that are wrongly classified in previous rounds A always have their weights decreased. B always have their weights increased. C may have their weights increased. D do not change their probability of being included in the training set. E None of the previous answers Question 24 Assume to have an ordinal attribute (which is not the class label) in your dataset, and that you decide to transform it into a numerical attribute, preserving the order. How will this affect the construction of a decision tree using the C4.5 algorithm? A Decision trees are only defined for nominal and numerical attributes, so the transformation is necessary to build the tree. B The resulting tree will be the same, but it will be much faster to produce it because with numerical attributes we do not have to check all combinations of values. C This transformation has no effect on decision trees D The resulting tree will be the same, but it will take a significantly longer time to produce it because numerical attributes have a larger number of possible splitting values. E None of the previous answers Corrected Question 25 Which of the following does not have an impact on the complexity of the k-NN algorithm during a classification process? A The number of records in the training data B The parameter k C The number of attributes D None of the other answers Consider the following training set TRAIN: a1 13 5 10 14 9 a2 1 5 3 4 2 a3 19 1 9 13 20 Class C1 C1 C1 C2 C2 Question 26 What are the distinct GINI impurities for all possible binary splits of attribute a1? (of course, do not consider splits generating empty nodes) A .3, .4 B .27, .4 C .27, .3, .4, .47 D .4, .47 E .3, .4, .47 F None of the previous answers Question 27 Which attribute would be chosen for the first split in a decision tree learning algorithm, using GINI and binary splits? A a1 B a1 or a2 C a2 or a3 D a1 or a3 E a2 F a3 G None of the previous answers Question 28 What are the distinct GINI impurities for all possible binary splits of attribute a3? (of course, do not consider splits generating empty nodes) A .27, .4 B .4, .47 C .27, .3, .4, .47 D .3, .4 E .3, .4, .47 F None of the previous answers Corrected Question 29 What are the distinct GINI impurities for all possible binary splits of attribute a2? (of course, do not consider splits generating empty nodes) A .27, .4 B .4, .47 C .3, .4, .47 D .27, .3, .4, .47 E .3, .4 F None of the previous answers Question 30 What are the distinct classification errors for all possible binary splits of attribute a3? (of course, do not consider splits generating empty nodes) A .2, .4 B 0, .2, .4 C 0, .2 D 0, .4 E .2, .4, .6 F None of the previous answers Question 31 What are the distinct classification errors for all possible binary splits of attribute a1? (of course, do not consider splits generating empty nodes) A 0, .2 B .4 C 0, .4 D .2, .4, .6 E 0, .2, .4 F None of the previous answers Question 32 What are the distinct classification errors for all possible binary splits of attribute a2? (of course, do not consider splits generating empty nodes) A 0, .4 B 0, .2 C 0, .2, .4 D .2, .4, .6 E .4 F None of the previous answers Question 33 Which attribute would be chosen for the first split in a decision tree learning algorithm, using classification errors and binary splits? A a1 B a2 or a3 C a3 D a1 or a2 E a1 or a3 F a2 G None of the previous answers Corrected Now consider the following training set TEST: a1 7 11 a2 2 4 a3 10 6 Class C1 C1 Question 34 What is the classification error of a 3-NN classifier with distance-based weighting trained on TRAIN and tested on TEST? (use Manhattan distance) A 3/6 B 2/6 C 5/6 D 4/6 E 6/6 F 0 G 1/6 H None of the other answers Question 35 What is the classification error of a 1-NN classifier trained on TRAIN and tested on TEST? (use Manhattan distance) A 6/6 B 5/6 C 1/6 D 4/6 E 0 F 2/6 G 3/6 H None of the other answers Question 36 What is the classification error of a 3-NN classifier with majority voting trained on TRAIN and tested on TEST? (use Manhattan distance) A 1/6 B 2/6 C 4/6 D 0 E 6/6 F 5/6 G 3/6 H None of the other answers Consider the following training set TRAIN: a1 13 10 5 14 9 a2 1 2 3 4 5 a3 19 10 9 13 8 Class C1 C1 C1 C2 C2 Corrected Question 37 Which attribute would be chosen for the first split in a decision tree learning algorithm, using GINI and binary splits? A a1 B a2 C a1 or a3 D a1 or a2 E a2 or a3 F a3 G None of the previous answers Question 38 What are the distinct GINI impurities for all possible binary splits of attribute a2? (of course, do not consider splits generating empty nodes) A .27, .4 B .4, .47 C .3, .4, .47 D .3, .4 E .27, .3, .4, .47 F None of the previous answers Question 39 What are the distinct GINI impurities for all possible binary splits of attribute a3? (of course, do not consider splits generating empty nodes) A .3, .4 B .3, .4, .47 C .27, .3, .4, .47 D .27, .4 E .4, .47 F None of the previous answers Question 40 What are the distinct GINI impurities for all possible binary splits of attribute a1? (of course, do not consider splits generating empty nodes) A .3, .4, .47 B .27, .4 C .3, .4 D .27, .3, .4, .47 E .4, .47 F None of the previous answers Question 41 What are the distinct classification errors for all possible binary splits of attribute a3? (of course, do not consider splits generating empty nodes) A 0, .4 B .2, .4 C .2, .4, .6 D 0, .2 E 0, .2, .4 F None of the previous answers Corrected Question 42 Which attribute would be chosen for the first split in a decision tree learning algorithm, using classification errors and binary splits? A a1 B a2 or a3 C a2 D a3 E a1 or a3 F a1 or a2 G None of the previous answers Question 43 What are the distinct classification errors for all possible binary splits of attribute a2? (of course, do not consider splits generating empty nodes) A 0, .2 B 0, .4 C .2, .4, .6 D .2 E 0, .2, .4 F None of the previous answers Question 44 What are the distinct classification errors for all possible binary splits of attribute a1? (of course, do not consider splits generating empty nodes) A 0, .4 B 0, .2 C 0, .2, .4 D .2, .4, .6 E .2, .4 F None of the previous answers Now consider the following training set TEST: a1 7 12 a2 2 4 a3 10 8 Class C1 C1 Question 45 What is the classification error of a 3-NN classifier with distance-based weighting trained on TRAIN and tested on TEST? (use Manhattan distance) A 3/6 B 2/6 C 0 D 5/6 E 6/6 F 1/6 G 4/6 H None of the other answers Corrected Question 46 What is the classification error of a 3-NN classifier with majority voting trained on TRAIN and tested on TEST? (use Manhattan distance) A 0 B 1/6 C 2/6 D 4/6 E 6/6 F 3/6 G 5/6 H None of the other answers Question 47 What is the classification error of a 1-NN classifier trained on TRAIN and tested on TEST? (use Manhattan distance) A 6/6 B 3/6 C 2/6 D 0 E 1/6 F 4/6 G 5/6 H None of the other answers Consider the following training set TRAIN: a1 13 10 5 14 9 a2 1 2 3 4 5 a3 19 10 9 13 8 Class C1 C1 C1 C2 C2 Question 48 What are the distinct GINI impurities for all possible binary splits of attribute a2? (of course, do not consider splits generating empty nodes) A .27, .4 B .3, .47 C .3, .4 D .4, .47 E 0, .27, .3, .4 F None of the previous answers Corrected Question 49 What are the distinct GINI impurities for all possible binary splits of attribute a3? (of course, do not consider splits generating empty nodes) A .4, .47 B .3, .4, .47 C 0, .27, .3, .4 D .27, .4 E .3, .4 F None of the previous answers Question 50 Which attribute would be chosen for the first split in a decision tree learning algorithm, using GINI and binary splits? A a1 or a3 B a1 C a2 D a1 or a2 E a2 or a3 F a3 G None of the previous answers Question 51 What are the distinct classification errors for all possible binary splits of attribute a1? (of course, do not consider splits generating empty nodes) A 0, .2 B 0, .2, .4 C .2, .4, .6 D 0, .4 E .2, .4 F None of the previous answers Question 52 What are the distinct classification errors for all possible binary splits of attribute a2? (of course, do not consider splits generating empty nodes) A 0, .4 B .2, .4 C .2, .4, .6 D 0, .2 E 0, .2, .4 F None of the previous answers Question 53 What are the distinct classification errors for all possible binary splits of attribute a3? (of course, do not consider splits generating empty nodes) A .2, .6 B 0, .2, .4 C 0, .4 D 0, .2 E .2, .4, .6 F None of the previous answers Corrected Question 54 Which attribute would be chosen for the first split in a decision tree learning algorithm, using classification errors and binary splits? A a1 or a2 B a1 C a3 D a2 or a3 E a1 or a3 F a2 G None of the previous answers Now consider the following training set TEST: a1 11 12 a2 4 5 a3 11 12 Class C1 C1 Question 55 What is the classification error of a 1-NN classifier trained on TRAIN and tested on TEST? (use Manhattan distance) A 1/6 B 5/6 C 0 D 2/6 E 4/6 F 3/6 G 6/6 H None of the other answers Question 56 What is the classification error of a 3-NN classifier with distance-based weighting trained on TRAIN and tested on TEST? (use Manhattan distance) A 6/6 B 1/6 C 0 D 5/6 E 2/6 F 4/6 G 3/6 H None of the other answers Corrected Question 57 What is the classification error of a 3-NN classifier with majority voting trained on TRAIN and tested on TEST? (use Manhattan distance) A 1/6 B 2/6 C 4/6 D 5/6 E 0 F 3/6 G 6/6 H None of the other answers