代做program、代写Python编程设计

日期：2025-02-28 09:12

General guidelines:

● This assignment is tentatively due at the end of week 8. But the final deadlines are as

posted on gradescope

● Given that this is a new class, welcome to provide feedback on any problems with the

spec!

● This assignment has 5 tasks, each worth 5% of your grade, for a total of 25%

● Assignments must be completed individually

● Assignment stubs are available here

Overall description

This assignment will require you to implement various fairness interventions to improve fairness

outcomes, while approximately maintaining accuracy, of a particular classifier.

In this assignment, we will predict default of credit card clients using the UCI “default” dataset:

https://archive.ics.uci.edu/dataset/350/default+of+credit+card+clients

Our label and sensitive attribute are as follows:

y = default payment next month (“y” == “1” in the dataset)

z = has graduate education (“EDUCATION” == “1”)

Files

The file autograder.py contains code to read the data and extract the label and sensitive

attribute. This is the same file the actual autograder uses, so should not be modified; it can be

run as-is to test your submission.

The file submission.py is for you to modify as needed. You should implement each of the

functions in this file. Existing implementations have already been provided to give you a sense

of how the assignment works. This is the only file you should upload to the autograder.

Tasks

You are required to complete the following 5 tasks, each worth 5 marks:

1. Make a solution that is as accurate as possible; you are allowed to use the sensitive

attribute.

2. Apply a dataset-based intervention to improve the fairness (demographic parity) of

your model; that is, you can modify the dataset however you’d like, but cannot alter any

other part of the modeling pipeline.

3. Apply a model-based intervention to improve the fairness of your model; here, the

sensitive attribute may be used during training but is not available at test time (i.e., it

cannot be a model feature).

4. Apply a post-processing intervention to improve the fairness of your model; here, you

are given the output probabilities of an existing classifier, along with the sensitive

attribute, and must generate output labels.

5. Apply any combination of interventions.

Baselines

The following baselines are provided for each task:

1. Just extend a trivial classifier (concatenation of all features) with the sensitive attribute.

2. Duplicate training instances with z=0

3. Assign higher instance weights to instances with z=0

4. Manually perturb the per-group thresholds a bit

Grading

Generally speaking, it should be possible to design interventions that satisfy demographic parity

(almost) exactly; partial grades are provided mostly for solutions that don’t quite make it. A

tentative grading criterion is as follows (same for Tasks 2-5; Task 1 will be graded more

trivially):

1 mark: Upload any valid solution (i.e., which doesn’t throw any error)

2 marks: Obtain a solution with roughly equal per-group TPRs

3 marks: Obtain a solution with roughly equal per-group TPRs without significantly reducing

performance compared to a trivial

1 model

4 marks: Obtain a solution with almost identical per-group TPRs without significantly reducing

performance compared to a trivial model

5 marks: Obtain the most accurate possible solution with almost identical per-group TPRs (i.e.,

accuracy almost the same as the topmost solution in the class)

The terms “roughly”, “significantly”, “almost identical”, and “most accurate possible” will be

defined in a few weeks, once several valid solutions have been uploaded.

1 Provided in the autograder.py file