- Deadline: 2008/11/17, 18:00 by e-mail
- (To): kameda[at]iit.tsukuba.ac.jp
- (Subject): PRML report2
- (Format): PDF

Form and test SVM classifier for the given dataset below.

**standard set**
Training: Set-A/Set-B, for #10, #100, and #1000.

Test: Set-A/Set-B, for #10, #100, and #1000 respectively.

Prepare SVM which can select linear / non-linear, with/without soft margin, some kernels (for non-linear SVM).

You don't need to write down your own code. You can introduce any library/software.

__You need to write down following topics:__

- Where to get (e.g. URL)
- Authors / Copyright holders
- Brief introduction

Form 48 kinds of SVM (shown later) by giving the training set and show the recognition result.

__Recognition results should show True-Positive, True-Negative, False-Positive, False-Negative.__

True / False : symbol assigned to the date

Positive / Negative : result shown by SVM.

For the 100 samples;

False | True | Sum | |

Negative | False-Negative | True-Negative | FN + TN = N = ? (expeected to be 100) |

Positive | False-Positive | True-Positive | FP + TP = P = ? (expected to be 100) |

Total | FN + FP = F = 100 | TN + TP = T = 100 | 200 |

**SVM (48 kinds)**

- Linear SVM without soft margin (3 kinds)
- #10
- #100
- #1000

- Linear SVM with soft margin , tring 5 different margins (5 x 3 = 15 kinds)
- #10
- #100
- #1000

- Non-linear SVM without soft margin (2 x 5 x 3 = 30 kinds)
- Kernel1 (polynomial) : tring 5 different parameters
- Kernel2 (Gaussian) : tring 5 different parameters
- #10
- #100
- #1000

- You can show the results of non-linear SVM with soft margin (commonly used style) too.

Test 48 kinds of SVM obtained in [2A-2] by the test set.

__Results should be shown in True-Positive, True-Negative, False-Positive, and False-Negative for them__

Write the discussion: The test results may be worse than the training ones. Is this true? Check it out and discuss it.

Form and test SVM classifier for the given dataset below .

**challange set**
Training: Set-A/Set-B, for #10, #100, and #1000.

Test: Set-A/Set-B, for #10, #100, and #1000 respectively.

Choose the best way of forming the best SVM (any of linear/non-linear, soft margin). You are asked to make only one (best) SVM this time.

Show the SVM you used and results (again in in True-Positive, True-Negative, False-Positive, and False-Negative) for #10, #100, and #1000.

Verify the SVM performance (of [2B-1]) with the test set. Show the results (again in in True-Positive, True-Negative, False-Positive, and False-Negative) for #10, #100, and #1000.

Write the discussion: Estimate the data set and imagine the distribution form (or formula behind the data sets).

Discuss the perfomance upper bound / limits of the SVM.

kameda[at]iit.tsukuba.ac.jp