Replies: 1 comment 8 replies
-
Thanks for sharing @lurosenb ! I personally am in favor of having this slightly separate to start with rather than in this repo, just to give it freedom to develop. If we add it here it's going to be part of the As you probably read from my review linked above in @lurosenb 's post I have some interest in this topic as well. Additionally, I've chatted with @matthklein in the past about such a project. Perhaps he's also interested (no pressure!). The only two things that are really important to me here are Your proposal indicates that you've thought about a) quite a bit, which is great! I definitely would love to hear what others think as well, of course. I think this sort of thing requires steering committee approval @MiroDudik @hildeweerts @adrinjalali but all opinions are welcome, of course! |
Beta Was this translation helpful? Give feedback.
-
Hi everyone!
I’m relatively new to the fairlearn community, but I’m excited to start contributing!
I’ve been working on moving an effort forward to introduce some unsupervised learning algorithms to the fairlearn ecosystem, specifically fair clustering algorithms. For an awesome overview of the space, Roman has a review here: review.
I put together a basic implementation/sample notebook for one of the fair clustering algorithms (fairlet k-means) on my fairlearn fork that replicates a well-known paper (the paper, my fork). However, after reaching out to Roman, Miro, and Hanna, it seems like it may be better to develop an incubation project (separate from fairlearn/fairlearn, perhaps fairlearn/fair-clustering or something) that centers on a specific application scenario for fair clustering.
There are more than a few potential applications for fair clustering algorithms, although pinpointing exact use cases can be tricky. To move this work forward initially, perhaps we could focus on opinion leaders and market segmentation, using samples culled specifically from social media, advertising and review based open source datasets (i.e. open source social graphs, reddit/twitter data and online reviews of films, music, etc.)
In these scenarios, features of interest can often be protected classes like gender, race or income. Market segmentation might mean targeted advertising to a specific group that provides them with unfair exposure to products, deals, or otherwise more attention paid to them as consumers. Identifying opinion leaders, which gives individuals undue influence, without balancing for the protected classes mentioned also appears fraught.
Determining a compelling application for the many existing methods of fair clustering presents a significant opportunity (and challenge). I could not find an instance of someone deploying one of these fair clustering approaches in a real-world-ish context (though please let me know if I missed something.)
I had a discussion this past week with Ana Stoica, one of Augustin Chaintreau’s students, who is looking into fair clustering in her own right. She has a background in graph theory and social networks, and showed me some interesting work on segmentation problems (Segmentation problems), fair spectral clustering (Guarantees for Spectral Clustering with Fairness Constraints) and even pointed me to yet another (very recent) paper on fair k-means clustering (Socially Fair k-Means Clustering). She’s also very excited about potentially contributing to a fair clustering effort through fairlearn, which is great.
I’d love to hear thoughts from the community on the above proposal! Specifically:
Thanks so much everyone!
Beta Was this translation helpful? Give feedback.
All reactions