CAREER: Participatory Design Methods for Algorithmic Systems

Lehigh University, Bethlehem PA

Investigators

Abstract

This research will develop methods for applying participatory design to the underlying components of algorithmic systems. Such systems incorporate increasingly complex algorithms from machine learning (ML), natural language processing (NLP), and related areas into user interactions, intending to achieve great benefits. However, cases of egregious errors, algorithmic bias, and other issues have revealed the shortcomings of relying primarily on quantitative performance metrics to inform design. A potential solution is to incorporate users into the design process. Although human-computer interaction research has developed numerous methods for user-centered design, many such approaches focus primarily at the interface level. This focus becomes problematic when a system's functionality is increasingly determined by ML models and algorithms. Furthermore, designing only for users of algorithmic systems can overlook other important relationships, such as the people whose data are being analyzed or those who may interpret the results. To address these issues, this research will develop participatory methods for human-centered design of algorithmic systems. These methods will be developed and tested by working closely with two non-profit organizations that already engage in data-intensive work but currently make limited use of algorithmic systems: AEquitas, which conducts legal analysis, and ProPublica, an investigative journalism newsroom. Unique challenges emerge when attempting to incorporate different people's relationships with a system into the design process. Existing participatory methods often use visual elements or manipulatives to represent interface components. The abstract mathematical formalisms of algorithmic systems, though, do not always lend themselves to such visual representations. This research will develop novel participatory design techniques to establish common ground between domain experts, who are less familiar with ML or NLP, and researchers, who are unfamiliar with the application domain. Doing so can leverage diverse participants' expertise and interpretations, thereby improving the fit between computational systems and existing practices. The participatory methods directly address underlying technical components, from feature selection, to model construction, to performance evaluation, to result interpretation. Furthermore, these methods will align those underlying technical aspects with current practices and lay understandings, increase the chance of catching and rectifying unanticipated egregious errors before they become problematic, and ensure the results are presented in a transparent and accountable manner. Finally, this process will inform the development of modules for classroom instruction, paired across STEM and social science courses, on how to incorporate human-centered concerns into designing algorithmic systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.