dashaasienga
commented
9 months ago @katcorr
I'm currently working on the Chapter 2 draft. It's still in progress, but one problem I'm facing is that my ggplot figures get cut off in the pdf:
It seems the boundaries don't appear in the pdf, even when I play around with the sizes of the plot. I can definitely look into this more, but before I spend too much time on it, I thought to ask you first in case you have a quick solution, perhaps from working with Clara or other students in the past :)
Thanks!
katcorr
@katcorr
We finally completed the STAT 495 paper so I have more time on my hands now to work on the Chapter 2 (I'm sorry for the delay with this, I just didn't expect the last week of classes to be as hectic as it's been :/)
P.S. I have no problem working on this during this and next week to compensate the time I wasn't able to spend on it last week. I only have 1 exam, so splitting my time before break between some thesis writing and studying for the exam isn't a problem for me :)
I'm using this issue to think through an outline for Chapter 2 just so I can be more organized in the writing process. I'd love any feedback regarding this!
The Standard ML Approach
Before diving into the mechanics of the Seldonian framework, I'm thinking of starting this section with an overview of the standard ML approach, discussing objective functions and how they are maximized/ minimized to arrive at a solution.
With this introduction, I'll discuss limitations of the approach from a technical standpoint, laying a foundation to understand why we observe algorithmic unfairness in the first place. I imagine going into detail about how objective functions are designed to maximize overall performance, which can have negative effects on minority demographic groups/ if the feature relationship is different for different demographic groups.
I could discuss potential remedies, which would be an organic segue into the next sub-section.
Overall, this section will be quite short and serve to motivate the topic of the chapter.
The Seldonian Framework
This section will heavily reference the original Science paper, the supplementary materials, and the website to provide a comprehensive overview of the algorithm, in particular, the mathematics and statistics behind it. I'd also like to highlight the distinction between Seldonian v Quasi-Seldonian algorithms. I will formally define the notation and walk through the key parts of the algorithm.
The aim is to serve as an introdution to the Seldonian Algorithm in theory.
Toy Example: A Seldonian Regression Algorithm
Finally, I'd end this section with the toy example we did in the Jupyter notebook. This section will include the code snippets to set up the problem as well as the results, compared to Ordinary Least Squares Regression. Finally, I'd include the experimentation results at the end.
The aim is to serve as an introdution to the Seldonian Algorithm in practice.
This will set the stage for the following chapters which will focus on a simulation study and an application to the COMPAS data set (synthesizing the Chapter 1 fairness definitions with the concept of the Seldonian Algorithm discussed in Chapter 2).
Question:
The Statistics thesis should be entirely reproducible. I'm wondering if that means I should include all the relevant code within the body of the thesis?
Let me know if you think I'm missing something, if I need to move some things around, or if a better organizational structure would make more sense!