I went to the library and asked for both Fengling Hu's thesis and Jessica Jeong's thesis. You were right! Fengling Hu's thesis was so beautifully written and really inspirational. I am still in awe and I definitely picked up some things that I hope I can emulate in my own thesis. Jessica Jeong's thesis was also really interesting but quite different from Fengling Hu's. It seemed more of a data analysis project with model comparisons, pretty similar to what my internship project and CS 247 final project looked like. I'd like to have a more theoretical approach for my thesis, however. What are your suggestions for striking a good balance between theoretical work, exposition, and data analysis? What would a successful thesis look like?
The article by the CS researchers at UMass helped identify possible directions for a simulation. As I continue to read and identify areas that interest me, it'd be helpful to know: what are the basics of a successful simulation study? This can help focus my thinking.
Literature Review
Besides the 2 previous theses, I spent most of my week reading a comprehensive paper Professor Spector shared on what is currently being done in the frontiers of fairness ML as well as looking into some of the papers referenced. I've included this paper in the articles section.
There was a lot of really useful information in this paper but I'll attempt to summarize the most important concepts. Statistical notions of fairness refer to fixing a small number of protected demographic groups and then asking for parity of some statistical measures across all these groups. Popular measures of fairness across groups include the positive classification rate, statistical parity, false positive and false negative rates, and equalized odds. As a next step, I’d like to understand the mathematical definitions behind these notions and I'm currently looking for helpful resources.
Something interesting to note is that these statistical notions of fairness by themselves do not on their own give meaningful guarantees to individuals or structured subgroups of the protected demographic groups. Instead, they give guarantees to “average” numbers of protected groups. Individual notions of fairness, on the other hand, ask for constraints that bind on specific pairs of individuals, rather than on a quantity that is averaged across groups. That is, similar individuals should be treated similarly along some defined similarity metrics. A caveat is that this method relies on strong assumptions regarding the data. It’s unclear whether this can be made practical although it is an important and ongoing research agenda. This is a gap in the literature that we could meaningfully contribute to. In trying to bridge the best of both worlds, researchers are trying to look into constraints that are practically implementable without the need of making strong assumptions about the data, but which nevertheless provide more meaningful guarantees to individuals and there are a few references the paper mentions regarding this new area of research. There are also a host of questions that this raises, for example: What function classes are reasonable? What “features” should be protected – those that are sensitive on their own or also those sensitive once you consider their intersection with other sensitive groups (ex. clothing styles when considering race and gender)?
There is also an issue that comes up on the dynamic of fairness. Often the composition of multiple fair components will not satisfy any fairness constraint at all. Similarly, the individual components of a fair system may appear to be unfair in isolation. For example, a job pipeline where a candidate must pass every stage may result in an unfair decision even if the components of the pipeline are statistically fair. Developing satisfying fairness definitions and richer frameworks that behave well under composition is an ongoing area of research.
Fair representation learning is a data debiasing technique that produces transformations (intermediate representations) of the original data that retain as much of the task-relevant information as possible while removing information about sensitive or protected attributes – essentially creating a construct space where protected attributes are statistically independent from other features. Another approach is repairing features to reduce pairwise dependence with the protected attribute using rank-preserving processes, likelihood-based approaches, and adversarial learning. I don't have enough understanding of this topic yet to determine if there is something meaningful and novel we could contribute to, but there are many familiar "statistical terms" like transformations and likelihood-based approaches, which makes me think that there is potential to contribute meaningfully to statistics research by diving deeper into this.
Finally, the literature is sparse outside classification ex. ranking, selection, and personalization. There is also an emerging line of work that considers causal notions of fairness [related to the UMass researches' work]. This is something we could explore.
Next Steps
It seems that the best next step is understanding the textbook mathematical and probabilistic definitions of various fairness metrics. It was a bit difficult to follow along with some of the technical explanations because I don't have a baseline understanding of the notation in the field. I'm currently looking for good resources for that.
As I work on that, I also need to begin narrowing down my area of focus. Right now, we have about 5-6 different areas that show promise and my biggest challenge at the moment is narrowing down the scope. Any advice on how to balance both learning as much about the entire field as I can while also identifying a specific area to dive deeper into? Out of the areas I've identified, is there any that you think would be more achievable?
General
I went to the library and asked for both Fengling Hu's thesis and Jessica Jeong's thesis. You were right! Fengling Hu's thesis was so beautifully written and really inspirational. I am still in awe and I definitely picked up some things that I hope I can emulate in my own thesis. Jessica Jeong's thesis was also really interesting but quite different from Fengling Hu's. It seemed more of a data analysis project with model comparisons, pretty similar to what my internship project and CS 247 final project looked like. I'd like to have a more theoretical approach for my thesis, however. What are your suggestions for striking a good balance between theoretical work, exposition, and data analysis? What would a successful thesis look like?
The article by the CS researchers at UMass helped identify possible directions for a simulation. As I continue to read and identify areas that interest me, it'd be helpful to know: what are the basics of a successful simulation study? This can help focus my thinking.
Literature Review
Besides the 2 previous theses, I spent most of my week reading a comprehensive paper Professor Spector shared on what is currently being done in the frontiers of fairness ML as well as looking into some of the papers referenced. I've included this paper in the articles section.
There was a lot of really useful information in this paper but I'll attempt to summarize the most important concepts. Statistical notions of fairness refer to fixing a small number of protected demographic groups and then asking for parity of some statistical measures across all these groups. Popular measures of fairness across groups include the positive classification rate, statistical parity, false positive and false negative rates, and equalized odds. As a next step, I’d like to understand the mathematical definitions behind these notions and I'm currently looking for helpful resources.
Something interesting to note is that these statistical notions of fairness by themselves do not on their own give meaningful guarantees to individuals or structured subgroups of the protected demographic groups. Instead, they give guarantees to “average” numbers of protected groups. Individual notions of fairness, on the other hand, ask for constraints that bind on specific pairs of individuals, rather than on a quantity that is averaged across groups. That is, similar individuals should be treated similarly along some defined similarity metrics. A caveat is that this method relies on strong assumptions regarding the data. It’s unclear whether this can be made practical although it is an important and ongoing research agenda. This is a gap in the literature that we could meaningfully contribute to. In trying to bridge the best of both worlds, researchers are trying to look into constraints that are practically implementable without the need of making strong assumptions about the data, but which nevertheless provide more meaningful guarantees to individuals and there are a few references the paper mentions regarding this new area of research. There are also a host of questions that this raises, for example: What function classes are reasonable? What “features” should be protected – those that are sensitive on their own or also those sensitive once you consider their intersection with other sensitive groups (ex. clothing styles when considering race and gender)?
There is also an issue that comes up on the dynamic of fairness. Often the composition of multiple fair components will not satisfy any fairness constraint at all. Similarly, the individual components of a fair system may appear to be unfair in isolation. For example, a job pipeline where a candidate must pass every stage may result in an unfair decision even if the components of the pipeline are statistically fair. Developing satisfying fairness definitions and richer frameworks that behave well under composition is an ongoing area of research.
Fair representation learning is a data debiasing technique that produces transformations (intermediate representations) of the original data that retain as much of the task-relevant information as possible while removing information about sensitive or protected attributes – essentially creating a construct space where protected attributes are statistically independent from other features. Another approach is repairing features to reduce pairwise dependence with the protected attribute using rank-preserving processes, likelihood-based approaches, and adversarial learning. I don't have enough understanding of this topic yet to determine if there is something meaningful and novel we could contribute to, but there are many familiar "statistical terms" like transformations and likelihood-based approaches, which makes me think that there is potential to contribute meaningfully to statistics research by diving deeper into this.
Finally, the literature is sparse outside classification ex. ranking, selection, and personalization. There is also an emerging line of work that considers causal notions of fairness [related to the UMass researches' work]. This is something we could explore.
Next Steps
It seems that the best next step is understanding the textbook mathematical and probabilistic definitions of various fairness metrics. It was a bit difficult to follow along with some of the technical explanations because I don't have a baseline understanding of the notation in the field. I'm currently looking for good resources for that.
As I work on that, I also need to begin narrowing down my area of focus. Right now, we have about 5-6 different areas that show promise and my biggest challenge at the moment is narrowing down the scope. Any advice on how to balance both learning as much about the entire field as I can while also identifying a specific area to dive deeper into? Out of the areas I've identified, is there any that you think would be more achievable?