fairlearn/community

fairlearn.org - everything related to the Fairlearn package. Please feel free to ask any kinds of questions!

Richard Edgar

@riedgar-ms

@hildeweerts , just merged your DCO instruction update. Thanks!

Hilde Weerts

@hildeweerts

Yay, thanks!

Kevin Robinson

@kevinrobinson

@marielledado hello! it's cool to see exploration of the dataset artifact, and then so many great questions about how such a datset might be used! it's been fun to follow along a little bit some of the things you and roman have been thinking through :) my main input would be that in order to really get into those substantive questions and tradeoffs about harms and benefits, there has to be a real deployment context. i'm happy to chat more sometime if there's ways to help collaborate!

Kevin Robinson

@kevinrobinson

hi folks! here's the mockup screenshot i shared for how sociotechnical "talking points" might look on a website

the rationale and thought process behind this is in this hackpad, Fairlearn: Sociotechnical "talking points": https://hackmd.io/nDiDafJ6TMKi2cYDHnujtA, please feel free to edit or comment if you like

and the "Fairlearn: Seed ideas for scenarios" hackpad: https://hackmd.io/gD8dwdPSRsqH3BxBt9terg?both has the discussion notes and votes that people cast for what scenarios would be best candidates for turning into a Jupyter notebook that illustrates how the Fairlearn Python library can reduce real harms for real people. thanks so much for everyone's thoughts and perspectives and help in moving this work forward together!

Kevin Robinson

@kevinrobinson

Screen Shot 2020-07-30 at 12.17.44 PM.png

Kevin Robinson

@kevinrobinson

also, the winning seed scenarios, with three votes each, were:

1. Identifying potential tax fraud
2. Debit card fraud investigation
4. Candidate screening
22. Predicting who needs special attention in healthcare

so i'll look into some of these a bit more and try to work through the guidelines in https://fairlearn.github.io/contributor_guide/contributing_example_notebooks.html and see what i can come up with

Roman Lutz

@romanlutz

@kevinrobinson this is fantastic! Apologies for having to drop early.
Anything I can help with? If you're planning to put notes, background research etc. into a hackpad I'm happy to read up a little myself as well.

Richard Edgar

@riedgar-ms

In a similar vein, would you like me to start thinking through 'talking points' (as you described) on some of the others, so we can also have the examples of cases where we think that technology isn't a good idea?

Kevin Robinson

@kevinrobinson

@romanlutz thanks! sure, i'll put something initial together on 8/3 and share it. i think it'll also be helpful to talk with you about how we can "deep dive" but also stay concise and actionable in a jupyter notebook format too. i'll take a stab at that in the draft

Kevin Robinson

@kevinrobinson

@riedgar-ms sure, that would be awesome! i'm happy to chat more or collaborate if that'd be helpful too

also, my vote is really to push us beyond the "should we build it or not" dichotomy. i think if we take a list of scenarios and say "our advice to data scientists in these situations is to not build that and quit their job" it is unlikely we are going to influence any data scientists in those situations :) so there might be a little bit of balancing of our team values and beliefs on one hand, and empathy with practitioners and the constraints they operate within. i'm hopeful by having to write small actionable "talking points" we can stay engaged with that tension productively, rather than just passing a judgment. i wrote a bit about that in the Fairlearn: Theory of change hackpad https://hackmd.io/EBh01XPtRZGHEg76oD1rGw so would love to hear your thoughts if you're interested. and also please just run with whatever you're thinking and try it out too, that would be awesome! it's always easier to discuss and move forward when there's something concrete :)

14 replies

Richard Edgar

@riedgar-ms

I have a PR out which splits up the requirements files for Fairlearn.... before I merge (likely tomorrow), I would welcome any feedback:
fairlearn/fairlearn#535

Roman Lutz

@romanlutz

@kevinrobinson sounds great! Looking forward to it.

Richard Edgar

@riedgar-ms

@kevinrobinson , I've started adding a few 'questions to ask' on some of the examples (not the ones which we are going to try to turn into notebooks).

Richard Edgar

@riedgar-ms

I also like the "Theory of Change" you've written up @kevinrobinson . One suggestion (probably goes under 'small bites' but feel free to incorporate where you think best): have simple tooling which can be incrementally introduced. We've already talk about how having a 'fairness module' isn't going to work because fairness isn't just a technical issue. But that also won't work because a single big module will either be used or it won't. If you have tooling that can go into lots of different places (or possibly better, be included as part of other tools), then you can encourage incremental adoption. And links back to a point we've used in closing: be humble, you're not going to address all the fairness issues at once.

Kevin Robinson

@kevinrobinson

@riedgar-ms these are awesome, thanks for adding in such thoughtful questions!! if you want to keep working on any of those, and especially to try to translate questions into "talking points" that prototype with simple tooling or actionable communication, i'd be super ll be excited to read through and talk more to try to work on that with you sometime too!

Kevin Robinson

@kevinrobinson

@riedgar-ms and +++ on the small tooling bit, i updated that copy, thanks!

Kevin Robinson

@kevinrobinson

for now i'm still working towards a fairlearn example notebook, and starting with "candidate screening." i also put together a graphic on different ways to engage in sociotechnical fairness work. i'll put that here, and a link in the "theory of change" hackpad as well, but would like to avoid discussing in the thursday meeting for now so we can stay focused on more concrete bits. but i'm trying to externalize and share my assumptions as i go too :)

the main point of the chart was to try to help me recognize as i was working that i kept drifting towards things that would be very hard for a practioner to influence, and to try to ground back to stay concrete and actionable

fwiw

Roman Lutz

@romanlutz

Interesting graphic! There's no meaning to columns, right?

Kevin Robinson

@kevinrobinson

nope!

Richard Edgar

@riedgar-ms

@kevinrobinson certainly happy to work some of those into talking points. Do you have a sample to follow, or would we be creating one :-)

Kevin Robinson

@kevinrobinson

@riedgar-ms i think all of this is creative work :) my initial thoughts on this are in https://hackmd.io/nDiDafJ6TMKi2cYDHnujtA, and if it's okay with you it'd be great to keep iterations of "talking points" in a separate place from the notes on "fairlearn example notebooks" (even if they evolve from the same set of notes on those initial seed scenarios)

2 replies

i'm happy to chat more if you want sometime too!

Kevin Robinson

@kevinrobinson

as an update, i worked a bit on candidate hiring, and think i have what can be a helpful discussion prompt for thursday. if folks are interested in previewing, here's a more specific case study: https://www.pymetrics.ai/case-studies/anz-case-study

two framing background bits are Rhagavan et al. 2018: https://arxiv.org/pdf/1906.09208.pdf and the examples and implementation notes for pymetrics/audit-ai: https://github.com/pymetrics/audit-ai/tree/master/examples

i think using the pymetrics example as a baseline can help make concrete or even validate some of the theory of change assumptions about fairlearn as well.

for thursday, i was thinking i could frame context for ~5m and then share some discussion prompts to kick things off, and see what happens for ~20m :) if folks have meta-thoughts on that approach or how to make thursday feel engaging and productive, please let me know!

Hilde Weerts

@hildeweerts

Awesome!

Kevin Robinson

@kevinrobinson

fwi i made another iteration on the chart above, Engaging with sociotechnical fairness work at different levels of "closure" i've been trying to use this to describe some of our work in the thursday meeting discussions, and while i don't want this to be a main part of the discussion today, i thought i'd preview it if folks are curious, and would love to present for 10m and get feedback in a week or two (like a mini-critique at a lab meeting) .

Engaging with sociotechnical fairness work at different levels of closure (v3).pdf

Andreas Mueller

@amueller

fyi I commented on the metrics thread, sorry if I'm late to the party

Kevin Robinson

@kevinrobinson

i emailed the "deep dive" list with links and notes from the meeting today. i can't just paste that here, because it loses all the links and formatting, but the main thing is

If you're interested in diving deeper on this, or collaborating on any part, please email or message in gitter.im/fairlearn/community anytime! This is the hackpad we worked through today: Fairlearn: Candidate screening

If you want the full email ping me or roman can help too since he owns that list :)

Kevin Robinson

@kevinrobinson

also if someone wants to help make a flowchart visual showing how a ANZ candidate goes through the hiring process, that would be super helpful too!

Sridharan Kamalakannan

@hidasri

Can anyone suggest best practice guidelines for selecting thresholds for the disparity metrics to determine if a model is biased towards a sensitive attribute or not?

Kevin Robinson

@kevinrobinson

hi @hidasri! if you can share more about the context for the prediction task, more folks might be able to help more! there's a bunch of folks in chat here who could help you work through thinking about this and how to pick different thresholds depending on how people might be harmed by a model's prediction errors

Varoon Mathur

@Varoon3

Hi @hidasri! Would love to spend some time thinking about your specific context if you wish to dive deeper into it - right off the bat in terms of metrics to determine disparity: those are highly contextual (i.e balancing how often those who require care are receiving it in and by themselves, with trying to understand how certain patient populations may compare with each other). Also more than happy to chat in any other setting if that helps!

LisaIbanez

@LisaIbanez

Hi @hidasri! Ditto what @kevinrobinson and @Varoon3 said...happy to help!

Roman Lutz

@romanlutz

In the last deep dive session we said we'd perhaps have one even the upcoming week (8/13), an hour earlier or an hour later than usual, i.e. 10am ET or 12pm ET (as opposed to 11am ET). I won't be able to make the later slot, and also have a conflict for part of the earlier one. What I'd suggest instead is to build on the work Kevin has done in the following ways and then check in at 10.30am ET for half an hour before the developer call to see if there's anything to coordinate. Let me know if you have thoughts!

notes from last time: https://hackmd.io/GMli82s7SxORABkabCgw8Q
todos:

read up on ANZ resources in there
think about the questions from section 5
think about how this could be turned into a notebook while highlighting real harms
look at the other 3 scenarios in https://hackmd.io/gD8dwdPSRsqH3BxBt9terg that got 3 votes. There are plenty of pointers in there for anyone to get started and research the context (laws, articles, papers) to find out more.

With any kind of progress on these fronts we could check in on Thursday and discuss issues, how to proceed, etc.

Richard Edgar

@riedgar-ms

@amueller @hildeweerts @MiroDudik (and any others who are interested), in the interests of making progress on:
fairlearn/fairlearn-proposals#12
Do you think it would be helpful to have a meeting?

LisaIbanez

@LisaIbanez

I could make different meeting times on Thurs.

Andreas Mueller

@amueller

@riedgar-ms I didn't have much time to look into it so far unfortunately and this week is quite busy already.

Hilde Weerts

@hildeweerts

I think a meeting would be very helpful! I only have time during the evenings the coming two weeks, but given the time difference that’s probably not an issue.

I’d love to get Andreas’s opinion on the metrics module, so if he’d be able to join if we do it next week I’d prefer that. Otherwise this week I’m available Wednesday or Thursday.

Richard Edgar

@riedgar-ms

@amueller how is next week looking for you?

MiroDudik

@MiroDudik

@riedgar-ms @hildeweerts @amueller : I'm on vacation next week, but I could possibly make an exception if that works for everybody else. I'll try to compose my thoughts a bit more re. most recent Andreas's and Richard's comments.

Roman Lutz

@romanlutz

@LisaIbanez looks like it's only the two of us (and Kevin said he may be able to join when we talked last Thursday). I spent a lot of my time since last week moving and haven't made much progress. I intend to change that today/tomorrow by following the steps I outlined above on 8/8. @LisaIbanez do you want to join in? The hackmds are all editable so we can add our thoughts and whatever we find :-) https://hackmd.io/GMli82s7SxORABkabCgw8Q

LisaIbanez

@LisaIbanez

@romanlutz Yeah, I can join in! So that means we are chatting at 7:30 PST or just making changes?

Where communities thrive

People

Repo info

Activity