Reducing Discrimination in Learning Algorithms for Social Good: My CURF Project


Did you know that higher-income neighborhoods on average have more access to smart mobility options than lower-income neighborhoods? After I evaluated the availability of Healthy Ride bikes throughout Pittsburgh using geospatial analysis, I discovered that only 5% of the bike stations were located in neighborhoods with poor housing conditions; the other 95% are located in census tracts that collectively had a median household income greater than $50,000 in 2015.

Location of Healthy Ride bike stations in Pittsburgh, PA colored by census tract. Size of point correlates to the capacity of each bike station. Hovering over a point reveals the census tract, the number of racks, and the 2015 median household income.

Smart mobility initiatives aim to provide affordable, on-demand, reliable modes of transportation such as bikes, electric scooters, or car sharing. The access to reliable transportation is vital to the stability and growth of a city connecting citizens to reliable jobs, medical facilities, education, and food. However, if the affordable, on-demand transportation is lacking presence in low-income neighborhoods, it excludes them from these amenities and opportunities.

Researchers have designed learning algorithms to tackle the classical resource allocation problem that planners and policymakers face when deciding where to allocate or redistribute resources, such as bikes or electric scooters. This pipeline unfortunately is vulnerable to biased predictions because the learning algorithms ultimately learn the biases in the data sets they are trained on before deployment. Biased predictions results in altering behaviors and making the data even more biased. This negative feedback loop is the sole motivation behind my CURF project this semester.

Related Works

There are very few academic contributions that have investigated or incorporated fairness in learning algorithms that aim to redistribute resources fairly. For example, [Yan and Howe, 2020] incorporate their own definition of fairness into their learning algorithm to fairly predict demand of bikes in Seattle. However, several works have investigated how to efficiently redistribute transportation resources to maximize usage and profit instead of equity across sensitive attributes such as race, income, or age [Luo et al., 2020; Lesmana et al., 2019; Rong et al., 2019; Cardona et al., 2017; Garcia et al., 2017]. A common technique used to determine where to redistribute bikes, or any limited resource, is to predict the demand of that resource at a given location and time [Yan and Howe, 2020]. Predicting this demand at a given space and time requires a data set of historical demand data for that given resource over time. This approach limits the model to learn biases inherent in the demand data, and it does not allow for exploring locations with no historical data [Mitchell, 2020].

Previous research has made significant contributions to the fair machine learning community. Researchers have been exploring how to mitigate discrimination and biases in learning algorithms by statistically defining fairness [Barocas et al., 2019; Hardt et al., 2016; Zemel et al., 2013; Dwork et al., 2011]. Some researchers have incorporated these defined notions of fairness into their frameworks to increase equity or opportunity among a protected attribute [Chen et al., 2020; Gölz et al., 2019; Donahue et al., 2020; Claure et al., 2020]. However, these works do not focus on resources in urban settings. The lack of fair learning algorithms for resource distribution in urban settings opened doors for further research.

CURF Project

My CURF project, under the advisement of Prof. Malihe Alikhani, aims to create a fair learning algorithm to provide insight as to where resources should be distributed to balance overall transportation equity across census tracts. This approach cannot and will not solely rely on historical demand data. Instead, we propose a fairness constraint on the demand prediction to achieve the two conflicting objectives of exploring uncharted areas and exploiting historically popular bike stations. The framework that we are developing takes into consideration attributes such as socioeconomic status, demographics, and location, that may not be represented in the demand data. Ultimately, the goal is to design a framework that can generalize to other resources such as dispatching law enforcement or dispatching food and water across remote villages.

Looking Ahead

Building fair and unbiased AI systems that can collaborate with people to create a more sustainable world is my research goal. I am currently waiting to hear back from several PhD programs to study human computer interaction, computer science, or social and engineering systems. My long-term career goal is to become a research scientist or professor with my own research lab focusing on human-AI interactions and fairness in AI systems. My experiences thus far have exposed me to some research, but I realized how much I still do not know, and how much room I have yet to grow. While I have taken steps towards meeting this goal, the Chancellor’s Undergraduate Research Fellowship will sharpen my research skills, help me to meet my potential as a researcher, and prepare me to be successful in a PhD program.

About Me

My name is Katelyn Morrison and I am a senior in the School of Computing and Information studying computer science and completing a sustainability certificate from the Swanson School of Engineering. Something unique about me is my love for biking and camping. I enjoy road biking around Pittsburgh and going bike camping on the San Juan Island off the coast of Washington. I cannot wait to pick up bike camping again once the pandemic subsides!


[Barocas et al., 2019] Solon Barocas, Moritz Hardt, and Arvind Narayanan. Fairness and Machine Learning., 2019.

[Cardona et al., 2017] Mateo Cardona, Juan Zuluga, and Diego Escobar. Analisis de la red de ciclo-rutas de manizales (colombia) a partir de criterios de accesibilidad territorial urbana y cobertura de estratos socioeconomicos. Revista Espacios, 38(28), 2017.

[Chen et al., 2020] Yifang Chen, Alex Cuellar, Haipeng Luo, Jignesh Modi, Heramb Nemlekar and Stefanos Nikolaidis. Fair contextual multi-armed bandits: Theory and experiments. In Conference on Uncertainty in Artificial Intelligence, pages 181-190, 2020.

[Claure et al., 2020] Houston Claure, Yifang Chen, Jignesh Modi, Malte Jung, and Stefanos Nikolaidis. Multi-armed bandits with fairness constraints for distributing resources to human teammates. In Proceedings of the 2020 ACM/IEEE International Conference on Human-Robot Interaction, pages 299-308, 2020.

[Donahue et al., 2020] Kate Donahue. and Jon Kleinberg. Fairness and utilization in allocating resources with uncertain demand. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, pages 658-668, 2020.

[Dwork et al., 2011] Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Rich Zemel. Fairness through awareness, 2011.

[Garcia et al., 2017] Zuluaga Garcia, Juan David, et al. Propuesta metodologica para el diagnostico y planifica- cion urbana de una red de ciclorrutas. Caso estudio: Manizales. PhD thesis, Universidad Nacional de Colombia- Sede Manizales, 2017.

[Gölz et al., 2019] Paul Gölz, Anson Kahng, and Ariel D. Procaccia. Paradoxes in fair machine learning. NeurIPS’19, 2019

[Hardt et al., 2016] Moritz Hardt, Eric Price, and Nathan Srebro. Equality of opportunity in supervised learning. CoRR, abs/1610.02413, 2016.

[Lesmana et al., 2019] Nixie S Lesmana, Xuan Zhang, and Xiaohui Bei. Balancing efficiency and fairness in on- demand ridesourcing. In Advances in Neural Information Processing Systems, pages 5309–5319, 2019.

[Luo et al., 2020] Man Luo, Bowen Du, Konstantin Klem- mer, Hongming Zhu, Hakan Ferhatosmanoglu, and Hongkai Wen. D3p: Data-driven demand prediction for fast expanding electric vehicle sharing systems. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 4(1):1–21, 2020.

[Mitchell, 2020] Margaret Mitchell. Bias in the Vision and Language of Artificial Intelligence. The Scottish Informatics and Computer Science Alliance. 2020.

[Rong et al., 2019] Huigui Rong, Qun Zhang, Xun Zhou, Hongbo Jiang, Da Cao, and Keqin Li. Tesla: A centralized taxi dispatching approach to optimizing revenue efficiency with global fairness. In KDD Workshop on Urban Computing (UrbComp ’20), 2019.

[Yan and Howe, 2020] An Yan and Bill Howe. Fairness- aware demand prediction for new mobility. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 1079–1087, 2020.

[Zemel et al., 2013] Rich Zemel, Yu Wu, Kevin Swersky, Toni Pitassi, and Cynthia Dwork. Learning fair representa- tions. In International Conference on Machine Learning, pages 325–333, 2013.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s