CSCE 970 Homework 3

Assigned: Tuesday, April 1

Due by 11:59:59 p.m. on Wednesday, April 23

Total number of points: 75

1. (65 points) Do your assigned part of the following. Of course, you should summarize your results in a detailed report with confidence intervals, just like in the previous homeworks.
   • Implement a bagging or boosting algorithm to run with your learning algorithm from homework 2 and report on how it affects error on the training and testing sets, including discussion on whether or not overtraining occurred.
     • Jiazheng Shi : boosting
     • Andrew Breiner : bagging
     • Chris Hammack : boosting
     • Nurzhan Ustemirov : bagging

   • Implement the weighted majority algorithm to combine your trained classifiers from homework 1 and homework 2 and report on how it affects error on the testing sets, including a comparison with each classifier on its own. You are to train each classifier on the original training sets, just like in the previous homeworks. Then train WM on an independent set (part of the test set). Then evaluate the combined classifier on the remaining part of the test set.
     • Cory Strope and Pooja Khati
     • Zhaohui Sun and Chang Wang

2. (10 points) State the approximate amount of time you spent on each problem.

• Extra credit item 1: Implement one of the nontrivial clustering algorithms from the book (NOT BSAS), e.g. MBSAS (Section 12.4), TTSAS (Section 12.5), the single link algorithm (Section 13.2.2), the complete link algorithm (Section 13.2.2), or the Isodata (hard k-means) algorithm (Section 14.5.1). For each algorithm you implement, vary the relevant parameters (e.g. in TTSAS, the thresholds and the order to which the vectors are presented to the algorithm). For each combination, graphically display the resultant clustering and comment on how ``good'' you think it is based on visual inspection.

  Run the algorithm on the given data sets. You may assume these data form compact clusters, so use point representatives in your algorithms if representatives are required in your choice of proximity measure. The data will be two-dimensional, so you may visually inspect the results of the clusterings. You may choose any appropriate proximity measure for your algorithm, and are only required to use one. No matter which algorithm you use, you should also report timing information. Given the amount of time taken and ``goodness'' of the best cluster, do you feel that your algorithm was a good choice for the given data sets? Describe all your results in a detailed report.

  The number of extra points you will receive depends on the complexity of the algorithm implemented, the thoroughness of your experiments, and the quality of your report. The maximum number of points is 20 per implementation.

• Extra credit item 2: Use the procedures in Sections 12.3.1 and 13.5 and Chapter 16 to determine the ``best'' number of clusters in the data (your choice of procedure will be determined by the clustering algorithm you use).

Back