Home Colloquium Series Research Group

MD4SG

Mechanism Design for Social Good Working Groups

Global Perspectives on Inequality


The MD4SG Inequality working group studies how optimization, incentive design, and learning can improve services for vulnerable populations, such as homeless individuals and smallholder farmers. We also study how algorithmic approaches can magnify existing inequalities, and how to design systems to guard against these effects. To address such questions, we draw on techniques from computer science, economics and operations research, as well as experience in social work and international development.

 

Spring 2019 Presentations

Presenter Title Abstract
Ben Otieno Biases in probabilistic machine learning risk scoring techniques

Inequality at times can be brought about unconsciously and in some cases consciously. In this talk, I will highlight some unconscious ways in which data and models bring about biases resulting to inequalities. The talk will focus on risk scoring that determine clientsí access to financial services and how the various risk scoring models financial institutions use act as a barrier to financial services access by the disadvantaged. The talk will also highlight the role data (missing data and qualitative data) plays in breeding inequality in the financial services sector. I will use small-scale farmers as a case study of the disadvantaged population.

Small scale farmers provide 70% of the food consumed by the world population, while the remaining 30% is provided by large scale farmers. The models financial institutions use to arrive at decisions whether to provide banking and insurance services are mostly modelled to fit the large scale farmers. This means the people who provide food to the majority of the population donít have good access to financial services, thanks to the risk models used and data relied on in the scoring. The talk will propose approaches that can be used to minimize the inequality.

Angela Zhou Policy Learning: Data-Driven Decision Making from Observational Data

In this talk, I'll focus on the question of policy evaluation and current research on leveraging machine learning and optimization for designing optimal treatment allocation rules. Although most of my work on optimal treatment allocation targets healthcare applications, in the context of the group on global inequality, I'll start out by drawing connections to the role that empirical evaluation has played in discussions regarding assessing interventions intended to improve services for vulnerable populations. One example is the discussion about comparing conditional vs. unconditional cash transfers, and how these different modalities require articulating welfare goals of interventions.

While observational data provides opportunities for causal inference in richer settings and from larger databases, it suffers from issues such as selection bias of individuals into treatments as well as the typical causal impossibility of observing outcomes for treatments not previously administered. I'll briefly talk about reduced-form approaches for policy evaluation that is more robust to violations of the typical unconfoundedness assumption. Lastly, these issues of confounding can also affect the evaluation of new proposed approaches on historical data, which we show can lead to "residual unfairness".


Fall 2018 Presentations

Presenter Title Abstract
Sera Linardi Jail: A Day in the Life

This talk draws from Sera's experience working with homeless and incarcerated individuals.

References:
Ana-Andreaa Stoica Algorithmic Glass Ceiling in Social Networks As social recommendations such as friend suggestions and people to follow become increasingly popular and influential on the growth of social media, we find that prominent social recommendation algorithms can exacerbate the under-representation of certain demographic groups at the top of the social hierarchy. To study this imbalance in online equal opportunities, we leverage new Instagram data and offer for the first time an analysis that studies the effect of gender, homophily and growth dynamics under social recommendations. Our mathematical analysis demonstrates the existence of an algorithmic glass ceiling that exhibits all the properties of the metaphorical social barrier that hinders groups like women or people of color from attaining equal representation. What raises concern is that the algorithmic effect is systematically larger than the glass ceiling generated by the spontaneous growth of social networks. We discuss ways to address this concern in future design and its pervasiveness under other algorithms.

Working Group Organizer

Bryan Wilder Ph.D. Student in Computer Science University of Southern California


Working Group Members

Rediet Abebe Ph.D. Student in Computer Science Cornell University
Jerry Anunrojwong Research Affiliate MIT; Chulalongkorn University
Milan Cvitkovic Ph.D. Student in Computing & Mathematical Sciences California Insitute of Technology
Mutembesa Daniel Researcher in Artificial Intelligence & Data Science Makerere University, Uganda
Bikram Datta Assistant Professor of Economics Indian Institute of Technology, Kanpur
Sera Linardi Associate Professor of Economics University of Pittsburgh
George Musumba Lecturer in Computer Science Dedan Kimathi University of Technology, Kenya
Benjamin Otieno Senior Lecturer in Computing & Informatics University of Nairobi, Kenya
Ryan Shi Ph.D. Student in Computer Science Carnegie Mellon University
Eric Sodomka Research Scientist in Economics & Computer Science Facebook Research
Ana-Andreea Stoica Ph.D. Student in Computer Science Columbia University
Angela Zhou Ph.D. Student in Operations Research Cornell Tech