CS 461: Machine Learning
Instructor: Kiri Wagstaff

CS 461 Homework 2

Due: Midnight, January 24, 2008

Please your answers to these questions in <yourlastname>-hw2.txt (or PDF).

Part 1: Decision Trees (40 points)

  1. Given this data set and partially built decision tree, calculate the impurity of nodes A though D using misclassification error (I = 1 - maxi (pi) where i ranges over all possible classes and pi is the probability of class i over just the items at the current node). Show your work. (10 points)

    weight (lbs) number_of_legs fur_color Class
  2. Now calculate impurity for nodes A through D using entropy (eqn. 9.3). Show your work. (10 points)

  3. Now consider splitting node A using the weight feature. Select a reasonable value for theta to split. Calculate the impurity (using entropy) that would be achieved if we split on this feature (eqn. 9.8). Show your work. (6 points)

  4. Given the decision tree shown in question 1 (no split on weight), what is its error rate on this test data set? (4 points)

    weight (lbs) number_of_legs fur_color Class
  5. Turn this tree into a set of three classification rules: (6 points)

  6. What is the difference between pre-pruning and post-pruning a decision tree? (4 points)

Part 2: Support Vector Machines (30 points)

  1. Why does a quadratic model generally require more training data than a linear model? (3 points)

  2. In gradient descent, what is the gradient and why do we want to descend it? (3 points)

  3. How can an SVM be applied to data that is not linearly separable? (4 points)

  4. Why do we seek to maximize the margin? (3 points)

  5. What are three common nonlinear kernel functions? (3 points)

  6. In the context of an SVM, what is the C parameter? (2 points)

  7. Given the following training data, the SMO algorithm produced the linear SVM weight vector shown.

    Is it a waltz?
    Tempo (bpm) Duration (sec) Year of composition Class (y_i)

    Weight vector:

    Tempo (bpm)Duration (sec)Year of composition

    with a bias value of 36.0399.

    Show how to classify a new item, x. (6 points)

    New item x
    Tempo (bpm) Duration (sec) Year of composition Class (y_i)
  8. Assume you have a problem with four classes. You want to build a single classifier that will output which class a new item is, but all you have is code for a binary SVM. How can you use this code to create a multiclass SVM classifier? (6 points)

Part 3: Evaluation (10 points)

  1. Why do we separate our data into training and test sets? That is, why not just have one data set, train the model, and then report its error rate on that data set? (4 points)

  2. What happens to a 95% confidence interval as N, the number of samples, increases? (3 points)

    Free answer: It becomes smaller (or tighter), because the increase in the number of samples increases our confidence that the mean is correctly estimated.

  3. Using McNemar's test, determine whether algorithm 1 and algorithm 2 have significantly different performance, at a significance level of 0.05 (95% confidence), given this contingency table: (3 points)


Part 4: Weka (20 points)

  1. (1 point) Run Weka and select "Explorer." Use the "Open file..." button to load haberman.arff. (Consult the Explorer Guide if you get stuck.) Set the class attribute to "class" and browse the class distribution plots for each attribute. Which attributes (if any) provide a good separation between the two classes?

  2. (4 points) Switch to the "Classify" tab and select "IBk" (click "Choose", then "Lazy", then "IBk"). This is the k-nearest neighbors algorithm. By default, k is set to 1. You can click on the "IBk -K 1 -W 0" line to change the K value (type in the "KNN" field). Under "Test options", select "Cross-validation" and set "Folds" to 10. Select "(Nom) class" from the pull-down menu to tell Weka which attribute you want to predict.

    Run IBk for k = 1 to 10 and report the accuracy obtained for each one.

    With 10-fold cross-validation, how large is the training set for each run? What accuracy do you get if you set k to this value, so that all neighbors are used when classifying new items?

  3. (7 points) Now let's see what the decision tree can do. Select "J48" ("Choose" -> "Trees" -> "J48"). Using the default options and 10-fold cross-validation, click "Start." What accuracy is obtained?

    By default, pruning is on ("unpruned" is False). Click on the "J48" line, set "unpruned" to True, and run it again. What accuracy do you get?

    What is the FP rate for Survival? for Died? Which one do you consider more important to minimize, and why?

    If you don't use cross-validation, just test on the training set, what is the accuracy rate for J48? How can you explain the difference in these results?

  4. (8 points) Now let's see what an SVM can do. Select "SMO" ("Choose" -> "functions" -> "SMO"). Using the default options and 10-fold cross-validation, click "Start." What accuracy is obtained?

    By default, SMO uses a linear kernel. Therefore, the learned model has weights for each feature, not for each support vector. What are the weights learned for each feature? What is the bias value?

    Click on the "SMO" line, set "useRBF" to True, and run it again (C=1 and gamma defaults to 0.01). How many support vectors are there? What accuracy do you get? Is there anything strange about the confusion matrix? Can you find a different value for gamma that gives a higher accuracy?

What to turn in

Upload this file to CSNS under "Homework 2":

  1. <yourlastname>-hw2.txt (or <yourlastname>-hw2.pdf if you prefer to submit in PDF format)

Extra credit

If you wish to tackle some extra credit, you may do so here. You can earn up to 20 points to be applied to any of your homework assignments (but not to exceed 100 on any assignment). To receive these points, you must get at least a 70% on the main part of Homework 2, so finish the regular assignment before moving on to this part.

A) 5 points: Download the NearestNeighbor.java solution from Homework 1 and modify it to output a confusion matrix summarizing the results of classifying the items in the test file.

B) 10 points: Download the NearestNeighbor.java solution from Homework 1 and modify it to perform k-Nearest-Neighbor classification. k should be an additional parameter that is specified on the command line.

You may combine both A) and B) in a single implementation.

C) 5 points: Do some research and find out: what is the different between the ID3 and C4.5 decision trees? What criterion do they use for picking a feature to split on? What else about these algorithms differs from our dicussion of decision trees in lecture? Use your own words; do not copy/paste quotes unless you put them in quotation marks and attribute them to their author. Put your answers in <yourlastname-dectrees.txt.

What to turn in

Upload these files to CSNS under "Homework 2: Extra Credit":

  1. NearestNeighbor.java
  2. <yourlastname>-extracredit.txt (or PDF): Text file describing which of the extra credit options you are attempting and submitting. If you included option B), indicate how to run your program from the command line (specifying a value for k).
  3. For A) and/or B): <yourlastname>-sampleoutput.txt (or PDF): Include sample output from running your program on the iris and haberman data files.
  4. For C): <yourlastname>-dectrees.txt (or PDF)