Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
    Niels Bantilan
    @cosmicBboy
    Welcome to the Themis-ml Gitter Channel!: The purpose of this room is to generally discuss topics/issues related to fairness-aware machine learning and/or the themis-ml package specifically. Feel free to start threads related (but not limited to):
    • datasets to integrate into the package
    • techniques from papers/blogposts that you'd like to see implemented
    • themis-ml feature requests
    Sara(Yuexi) Li
    @YuexiSC
    Hi Niels!
    Sara(Yuexi) Li
    @YuexiSC
    I am a graduate student majoring business analytics at Santa Clara university. This program will finish at the end of this year. I have software engineering background in my bachelor degree, but its kind of testing-track. I watched your presentation on New York Open stats meetup. Thats really impressive. Data ethic is a really hot topic and many people around silicon valley are doing this. I read your paper so many times and really amazed by the behind logic dealing with discriminations. The whole processes is quite consistency and connected. I see that actually each module works just need to fill up with different methods, upgrade the time complexity and probably show a better visualization.
    Sara(Yuexi) Li
    @YuexiSC
    Few of my thoughts: So far as I am dealing with the whole project, the most time and effort were spent on getting know the logic of why certain method, how this method works and wait for the result. So I am thinking except to complete the modules, maybe it would be better to deliver in a more intuitive way. Like creating a final dashboard show how each method improve the bias in the dataset. I mean, automate the whole process and then people wont need to spent much time on look the paper over and over again and grasp the whole logic. Nowadays people seems embrace more on packages which are more simple to use, more easy to get the results and more integrated ones. And many trending packages seems like so. I know this might sound naive but just my personal opinion. ;)
    Sara(Yuexi) Li
    @YuexiSC
    To be honest, I want to contribute because I will face job hunting few months later, so I want to prove that I am been part of some project thats demanding. Having said that, I deeply know how it feels if been part of something I dont like. So thats why I turned to you and want to discuss to you is there anything I can help with.
    Niels Bantilan
    @cosmicBboy
    Hi Sara, thanks for sharing this!
    it's helpful to know more of your background, and thanks for the feedback on the project :)

    So I am thinking except to complete the modules, maybe it would be better to deliver in a more intuitive way.

    totally, I initially conceived themis-ml as simply an extension of the sklearn interface to be able to handle the abstractions required to do fairness-aware ML, I think the delivery in the form of visualizations/dashboard would be super useful

    we can probably chat about this more in detail, I think the main challenge for these kinds of more automated interfaces is the problem of over-generalizing an interface before knowing where the bulk of the use cases lie.

    in this way there's a trade-off between lower-level procedures (as you mention grasping the logic of the underlying methods) and having an interface that abstracts all of those things away (much like the fit, predict, etc. methods in sklearn do), except at an even higher level (such that the user doesn't really need to know the underlying logic of the methods)

    Niels Bantilan
    @cosmicBboy
    just brain-storming here, but I can think of two approaches:
    1. themis-ml as only providing the sklearn-like high-level interface to lower-level fairness-aware ML methods, and having another library altogether be responsible for visualizing the results of fairML methods.
    2. themis-ml as providing an additional interface to gather insights from a particular set of models, e.g. seeing how a fairML method decreases social bias with respect to some naive method
    I don't have a strong opinion about this right now, but it sounds like from what you've said that you're more interested in the insights layer to be able to present the results of themis-ml models in a more digestible way... would that be correct for me to say?