Computer scienceData scienceMachine learningEnsemble learning

Adaboost: an illustrated walk-through

Updating the weights

Report a typo

Suppose the team of soda developers decided to introduce a line of energy drinks. One of their first items is the Mystery drink. They asked a group of energy drink connoisseurs to test the first batch and complete a short survey about their preferences. The responses look like this:

+----+--------------+------------------+------------------------+--------+
|    | zero sugar   | blue coloring?   |   caffeine content(mg) | take   |
|----+--------------+------------------+------------------------+--------|
|  0 | yes          | no               |                    400 | yes    |
|  1 | yes          | yes              |                    125 | yes    |
|  2 | no           | no               |                     85 | no     |
|  3 | yes          | yes              |                    300 | yes    |
|  4 | no           | yes              |                    275 | yes    |
|  5 | yes          | no               |                    200 | no     |
|  6 | no           | no               |                     85 | no     |
+----+--------------+------------------+------------------------+--------+

The following stumps were obtained:

3 sample decision tree stumps based on the features

Initialize the weights for this dataset. Calculate the amount of say for the caffeine content split. Update the weights based on the caffeine content split. Normalize the resulting weights. The answer should contain 2 space-separated values: the new weight for the correctly classified sample and the new weight for the incorrectly classified sample, rounded up to the second decimal.

As a reminder, the amount of say is calculated as:

AOS=12ln⁡1−total errortotal error\text{AOS} = \frac{1}{2} \ln \frac{1- \text{total error}}{\text{total error}}

Note: the resulting normalized weights should sum up to 1, but due to the rounding, they won't add up to 1 in this task.

Enter a short text

Create a free account to access the full topic