Question

5. a. Suppose that the target variable in a classification task is continuous numeric, can you apply decision trees directly to it? How can you work around this? 5 points b. Models A and B has an error rate of 15% and 25% respectively. Further, the AUC for models A and B respectively is 65% and 70%. Which of these models is better? 5 points c. Would you prefer more to less false positive or false negative in each following cases? Justify your reasoning. 5 points i) Court trial of a robbery case. Positive class is guilty ii) Testing for a deadly disease. Positive class is positive test result. d. Discuss the advantage and disadvantage of using a small value versus a large value of k in k-nearest neighbor classification algorithm. 8 points

Answer 1

5a. If the target variable in a classification task is continuous numeric, we cannot apply decision trees directly to it.The target variables in a classification should be discrete. We need to convert continuous variable to discrete variables to apply decision trees on it. For example age may contain continous values like 25,45,67,23,50 etc. This can be converted to discrete variables like 1 represents age <40, 2 represents age>41 and age<60 and 3 represents a>61.

5 b. The first models with A 15% and B 25% is better as the error rates are less in this model compared to the other one.

5 c.1. Prefer less false positive. This is the case when a person is found guilty when he is not actually guilty.

5 c 2. Prefer less false negative. False negative in this case means the test case shows negative when actually the disease is present.

5 d. In K-nearest neighbour classification algorithm, k decides the number of partitions. To select the K that’s right for your data, we run the KNN algorithm several times with different values of K and choose the K that reduces the number of errors we encounter while maintaining the algorithm’s ability to accurately make predictions when it’s given data it hasn’t seen before.

· As we decrease the value of K to 1, our predictions become less stable.

· Inversely, as we increase the value of K, our predictions become more stable due to majority voting or averaging, and thus, more likely to make more accurate predictions (up to a certain point). Predictions with a high value of K beyond a certain point will be unstable or erroneous. In most cases, the optimum K-value lies between 5 and 30.