Abstract

Participatory Design (PD) offers a methodological framework for including affected communities in technological design. However, PD does not provide a theoretical model of how to ensure lasting social change with the produced technology. As a result, projects that include PD may not have long-term viability. We propose the use of Transformative Social Innovation (TSI) theory to guide PD-informed machine learning (ML) projects, and specifically “ML for social good” projects, towards technology that has the potential to create sustainable social change. In this position paper, we discuss how TSI, as a theoretical lens, intersects with and extends PD.

Keywords

social work, transformative social innovation theory

Introduction and Motivation

Machine Learning (ML) has demonstrated a clear interest in developing technology that makes lasting contributions to the betterment of society. One critique, particularly of earlier work in this space, though, is that the term “social good” can be poorly defined. Definitions often focus on concrete but narrow measures, such as a specific mathematical definition of “fairness,” or on a static but potentially uninformative outcome, such as predictive accuracy on imperfect historical data [4]. Such definitions are ineffective in many ways; here we will focus on one we find particularly important. These definitions that constitute the “social good” of a given piece of technology do not lay out the processes by which this technology might actually meet its ultimate goal of leading to sustainable social change if implemented in the real world.

A lack of interest in this up-front thinking about the long-term goals can lead to machine learning for social good (ML4SG) technologies that become one-offs and/or are unsustainable [8].One more specific case in which such upfront thinking is critical is to ensure that ML4SG technologies will be informed by, and center the needs of, those most impacted by the technology. A lack of consideration of this in turn leads to uninformed assumptions about what might benefit them in sustainable ways. Noting this, scholars have begun to use Participatory Design (PD) methods in ML4SG. PD has seen important successes in this area. For example, Lee et al. [2019] develop WeBuildAI, a participatory framework, which is then applied to the problem of equitably distributing food donations. Similar to early social innovation practices, best practices that have informed PD in this space have emerged primarily from analysis of case studies [7]. However, PD is an approach to improve upon technology design and development. It is not, and was never intended to be, something that could offer direction on how to reflect upon the long-term goals to transform institutions and empower citizens by embedding change in the fabric of social systems. As a methodological framework, PD therefore does not provide theoretical tools that help guide us towards sustainable social good within complex sociotechnical systems. Consequently, while PD is a critical method for ML4SG projects (and it allows them to aim for informed impact), its use could be extended by placing it within a theory for sustainable, positive social change in real-world sociotechnical systems.

We propose the adoption of Transformative Social Innovation theory (TSI) [5] as such a theory, within which PD frameworks fit cleanly. TSI offers data scientists (1) the opportunity to make explicit the theoretical lens that guides the intentions of their innovations for social good; (2) assistance in mapping the path from “tech for social good” to “good tech for sustainable social change”; and (3) a lens from which to consider success metrics, based on a theory of change. For example, If adopted by WeBuildAI, TSI theory could help researchers lay out a path to placing WeBuildAI into the complex web of people, regulations, and organizations that make up food donation services in a way that sustainable change could occur. In this position paper, we introduce TSI and then discuss early work on applying it to the problem of allocating services to youth in foster care.

TSI defines a transformative social innovation as one that 1) inspires sustainable changes in institutional agendas and 2) makes significant impacts on routine work, with 3) the goal of contributing to broad social outcomes, which 4) require the use of design processes that empower affected groups of people [2]. Note that the last piece of this definition is essentially PD, and thus TSI and PD have important overlaps. TSI theorizes that social innovation that changes society for the better occurs primarily through participatory actions between people and the organizational systems where the innovation will take place, as well as the impacts of the macro-social systems (e.g. social class structures) they are embedded within. Notably, then, TSI puts much less focus on the quality of the technological innovation. TSI provides a new lens to think about ML4SG projects and how to improve their sustainability and usefulness to society, because it largely focuses on who designs the technology and what comes after the technology has been built.

TSI leverages a variety of concepts to help teams theorize a path by which their ML4SG technology can lead to lasting social change. First, TSI formalizes the notion of a socio-material context into which the technology will be placed. A socio-material context is defined as the people, materials and technologies, institutional rules, and the interactions between these things [5]. TSI considers this context within the broader context of social values, politics, and other macro issues that influence change goals. Structurization, i.e., formalizing and embedding change in a system, occurs through affecting the system traditions through collaborative innovation. The outcome of social innovation is assessed by the degree to which social innovation is embedded in the new system. In the rest of the paper, we focus on two collections of critical concepts in TSI. These concepts rely on, but do not focus on, concepts of socio-material context and structurization.

First, embedding social change so it is sustainable is theorized to be affected by change contexts referred to as four “shades of change” [1]:

Social innovation is a change of relationships in the socio-material context that leads to new ways of “doing, organizing, framing, and knowing” [5]. In this sense, an emphasis is placed on the word “social” in social innovation, in that the key to innovating relies on changing the social system instead of introducing a product, such as new technology.
System innovation is a change to the structure or operation of a social or physical system, e.g. by disrupting or replacing it.
Game changers are macro events or trends that compel a need for an innovation.
Narratives of change are the discourses around the target of change, and focus on creating counter-narratives to combat narratives about why things are organized in a current system.

Second, TSI references social enablers of change - things that can help push forward the desired, positive changes. For example, social enablers include a “third sector” or neutral ground that can bring people together to enact participatory design [2].

An Example Application of TSI for ML4GOOD

We will take as an example application of TSI our current work to improve services for youth in foster care. The transformative social innovation in this project is a machine learning model to help child welfare caseworkers make service decisions that might most benefit older foster youth (ages 14-21), while doing so in an equitable manner.

The system innovation in this project is a reconceptualization of the ways that data are used in the child welfare system.Service allocation currently relies on caseworker decision making, which can be inconsistent, biased, or rely on limited practice experience due to high workforce turnover [3]. The current project relies on consultation with youth who have experience in foster care and seeks to develop ways to help present the youth’s perspective on service allocation to case workers. As a result, the system innovation will result in shifts in power related to decision-making (from case-workers to youth in the foster care system), which has broad system implications. The project conception relies on the intersection of multiple macro game changers, e.g. improved computing power and technological innovation in society.

Some of the narratives of change that influence our work and constrain our social innovation goals include popular cultural discourse and scholarly research about the negative impacts of algorithms on people impacted by them, and the mechanistic versus humanistic productions of machine learning. Finally, an example of a shade of change includes a third sector, created via our research team and its established relationships with youth, case workers, administrators, and government officials. This third sector helps address existent power differentials.

Conclusion

TSI offers a way for ML4SG researchers to formalize their theories of change; that is, their notion of how the tool(s) they are building will create positive, sustainable social change. TSI can help guide researchers in framing problems, designing technical innovations, and evaluating their impact in a way that drives real and lasting changes for social good. Further, it does so in a ways that are consistent with current pushes towards PD, creating a better means of applying this already useful approach to design.

References

Flor Avelino, Julia M Wittmayer, Bonno Pel, Paul Weaver, Adina Dumitru, Alex Haxeltine, René Kemp, Michael S Jørgensen, Tom Bauler, Saskia Ruijsink, et al. 2019. Transformative social innovation and (dis) empowerment. Technological Forecasting and Social Change 145 (2019), 195–206.
Karina Castro-Arce and Frank Vanclay. 2020. Transformative social innovation for sustainable rural development: An analytical framework to assist community-based initiatives. Journal of Rural Studies 74 (2020), 45–54.
Rebecca Jean Featherston, Aron Shlonsky, Courtney Lewis, My-Linh Luong, Laura E Downie, Adam P Vogel, Catherine Granger, Bridget Hamilton, and Karyn Galvin. 2019. Interventions to Mitigate Bias in Social Work Decision-Making: A Systematic Review. Research on Social Work Practice 29, 7 (2019), 741–752.
Luciano Floridi, Josh Cowls, Thomas C King, and Mariarosaria Taddeo. 2020. How to Design AI for Social Good: Seven Essential Factors. Sci Eng Ethics (2020).
Alex Haxeltine, Flor Avelino, Bonno Pel, Adina Dumitru, René Kemp, Noel Longhurst, Jason Chilvers, and Julia M Wittmayer. 2016. A framework for transformative social innovation. TRANSIT working paper 5 (2016), 2–1.
Min Kyung Lee, Daniel Kusbit, Anson Kahng, Ji Tae Kim, Xinran Yuan, Allissa Chan, Daniel See, Ritesh Noothigattu, Siheon Lee, Alexandros Psomas, and Ariel D. Procaccia. 2019. WeBuildAI: Participatory Framework for Algorithmic Governance. Proceedings of the ACM on Human-Computer Interaction 3, CSCW (Nov. 2019), 181:1–181:35.
Clay Spinuzzi. 2005. The methodology of participatory design. Technical communication 52, 2 (2005), 163–174.
Kentaro Toyama. 2015. Geek heresy: Rescuing social change from the cult of technology. PublicAffairs.

Transformative Social Innovation as a Lens for ML for Good