Why we participate
Credit scoring and other black box models in finance shape our life tremendously. They are the reason one might not receive a credit card or a loan or might not be considered a business partner. Since both, models and data are often owned by private companies, the ones who are affected by them have no insight and no handle to find biases within these. We want to change that. We deeply believe that each powerful tool needs to be controlled thoroughly and that people must have the opportunity to understand and dispute its negative impacts.
What we offer
We want to empower people to understand black box models and to find their underlying biases. Therefore we offer a public dataset example so that users can build trust to our approach.
How it works
We collect data where we know the in- and output. The output is generated by the black box model. Since we don't have access to this model (and we never will have, due to it's being under private ownership) we train our own model, which shall approximate the black box model as good as possible. Once we are close enough to the black box model we use advanced methods from game theory developed by Nobel laureate Adam Shapely to explain it's outcomes. We can therefore explain how important each feature is and how changes in the inputs affects changes in the output, so that one can find out whether one would still be declined a loan if having a different gender, race or a higher education degree.
What is the road ahead
Our vision doesn't stop using publicly available data. We believe that true ownership of data is to realize what can be done with it and how it compares to personal data of others. Therefore we want to offer the chance to donate anonymous data so that we can share insights on more models.
What it does
We want to explain why machine learning models behave the way they do and show everybody affected by them what influence small changes in the features can have to the output produced Therefore we used some public data (the adult dataset) to train a model and use a game theoretical approach to explain its outcome
How We built it
We initially built it using Python and Flask however, we found that using Streamlit gives us more flexibility and time to develop and roll out our App was quicker with Streamlit. So we built our initial prototype with Streamlit and deployed it with Microsoft Azure
Challenges We ran into
The speed of running our models & our App wasn't too great, so we looked at various ways that we can improve the performance of our App
Accomplishments that we're proud of
We built our custom dashboard after overcoming many technical obstacles and difficulties along the way We also learnt how to build a super-cool dashboard using Streamlit which worked really well for us #thankyoustreamlit
What We learned
We learned a lot about working remotely and between different timezones....and also how easy it was to build a dashboard with Streamlit
What's next for Bias Analyzer
Bias Analyzer still has a long way to go We have more datasets that we can add to improve our models. Users can donate datasets to us so we can improve our models as well at the outputs that we produce The performance of our dashboard still isn't that great, so we've got a bit of work in terms of enhancing our performance with larger volumes of data We can also add in more metadata for our datasets
Log in or sign up for Devpost to join the conversation.