Experimenting with different IBM AIF360 classifiers/algorithims on the COMPAS dataset (included in aif360.datasets).
More information on the COMPAS dataset and examples are found in the PDF, located here.
The functions file contains all necessary functions as well as docstrings with how they work. The example file shows usage of these functions, along with sources and other information about this research.
A summary of objectives and learned/completed outcomes is available here.
To-do list with objectives that could be completed in other semesters:
- Add a way for eta to be a bounded value from 0 to eta_bound+1 and plot all values of eta in this range to the upper-bound. This couldn't be completed due to computational cost (i.e: my computer had boot-failed several times after attempting to plot and compute the fairness metrics for eta within a range, thus the function will only take one value of eta).
- Look at different datasets and how their metrics are graded. This could be done by using the functions and tweaking the "fetch_input" function.
Here are some key values of ETA to test:
- 0
- 250
- 500
- 1000
The difference in fairness metrics for Prejudice Remover is seen at higher values of eta. This is most noticeable if you compare the fairness metrics from ETA=0 and ETA=250.