04-23, 16:35–17:05 (Europe/Berlin), A1
This presentation addresses the rare use of machine learning fairness metrics in domains with indirect human impact, e.g., automotive engineering. We briefly map out the space of use cases to examine the necessity, potential benefits, and challenges of applying fairness-related techniques. The main focus then lies on proposing solutions for overcoming identified hurdles, especially regarding the application in unstructured data domains, such as image and audio recognition and large text document analysis. Our approach includes strategies for detecting key subgroups and providing clear explanations for model failures. We also highlight two open-source tools, Sliceguard and Spotlight, for practical implementation.
Fairness Metrics are already widely used to avoid unwanted bias in machine learning models. However, although fairness is a hot topic, it is primarily used in domains where the models' interface and influence on humans are obvious. In other domains with a less obvious connection between model decisions and their impact on human beings, they are rarely seen (e.g., automotive engineering applications, etc.). This poses three questions:
- In those domains, is it really unnecessary to use fairness techniques, or is their absence endangering individuals in a less obvious way? (necessity)
- Even if a use case does not need fairness techniques, wouldn't the use cases still benefit from a look through the "Fairness lens" and the connected methods and tools? (benefit)
- Besides having less strong implications for using fairness metrics, what obstacles keep people from using them, and how can we mitigate them? (obstacles and solutions)
To answer these questions, our presentation will first briefly compare five prototypical engineering use cases and categorize them according to the above criteria (necessity, benefit, obstacles). This first part mainly aims to map out the space of machine learning use cases in the engineering domain and suggest possible reasons why fairness-related techniques are not applied in those areas.
We will then mainly focus on further analyzing those obstacles and providing solutions to omit them. Here, the main focus will be expanding the application of fairness-based model evaluation to unstructured data domains. Typical use cases in this category go from image and audio recognition to LLM applications with large text documents. We will provide a brief theoretical overview of strategies to make fairness metric application suitable and then go through a concrete example down to the implementation level. For that, we will touch on important subjects, such as detecting meaningful subgroups in unstructured data, extracting easy-to-grasp explanations for model failures, and interactive analysis of model predictions. This section will also feature two open-source tools to address these challenges: Sliceguard and Spotlight.
Intermediate
Expected audience expertise: Python –Novice
Public link to supporting material, e.g. videos, Github, etc. – Abstract as a tweet (X) or toot (Mastodon) –Exploring the need for fairness in machine learning in indirect human impact areas, proposing solutions for challenges in unstructured data.
I'm a seasoned AI professional with an additional background in software engineering and web development. Having participated in and led quite a few machine learning-based projects in the engineering domain, I've worked on various ML problems, ranging from ML on images and 3D data to audio and time series analysis. My expertise in software engineering makes it easy for me to bring ML solutions to production. Currently, my focus is on helping people build effective ML models, from planning out projects to creating performant models and productionizing them. I'm particularly passionate about data curation, and I'm excited to be a part of the team building Renumics Spotlight, a free data curation tool.