Abstract
New technologies and the availability of geospatial data have drawn attention to spatio-temporal biases present in society. For example: the COVID-19 pandemic highlighted disparities in the availability of broadband service and its role in the digital divide; the environmental justice movement in the United States has raised awareness to health implications for minority populations stemming from historical redlining practices; and studies have found varying quality and coverage in the collection and sharing of open-source geospatial data. Despite the extensive literature on machine learning (ML) fairness, few algorithmic strategies have been proposed to mitigate such biases. In this paper we highlight the unique challenges for quantifying and addressing spatio-temporal biases, through the lens of use cases presented in the scientific literature and media. We envision a roadmap of ML strategies that need to be developed or adapted to quantify and overcome these challenges -- including transfer learning, active learning, and reinforcement learning techniques. Further, we discuss the potential role of ML in providing guidance to policy makers on issues related to spatial fairness.
The Problem: Beyond I.I.D. Assumptions
Most machine learning (ML) fairness research assumes that data points are "independent and identically distributed" (i.i.d.). However, geospatial data—information tied to specific locations and times—violates this assumption. Nearby locations are often similar (spatial autocorrelation), and historical decisions, such as redlining or infrastructure placement, create deep-seated biases that standard algorithms fail to detect.
This paper addresses a critical gap: while ML is increasingly used to decide where to deploy 5G, send ride-share drivers, or distribute vaccines, we lack a rigorous framework to ensure these decisions are fair across different neighborhoods and time periods. Without such a framework, ML risks creating "black holes"—underserved areas that disappear from datasets because they lack the sensors, smartphones, or infrastructure to be "seen" by the model.
A Conceptual Framework for Spatio-Temporal Bias
Rather than introducing a single new algorithm, the authors provide a conceptual framework and research roadmap. They categorize five primary sources of spatio-temporal bias and map them to potential mitigation strategies:
- Geographical Differences: Natural variations (e.g., flood zones) that disproportionately affect minority communities.
- Mental Maps: Human prejudices that influence where services (like Uber or TaskRabbit) are offered or used.
- Spatial Clustering: The tendency for demographic groups to live in specific clusters, leading to uneven access to 5G or broadband.
- Spatio-temporal Data Glitches: Non-random missing data, such as the "pothole app" problem where data is only collected from citizens with high-end smartphones.
- Feedback Loops: When models prioritize areas with high "demand," but that demand is actually just a reflection of where services were historically available.
Proposed Mitigation Roadmap
The proposed roadmap suggests adapting several advanced ML techniques to address these biases:
- Active Learning: To identify and prioritize data collection in data-poor regions.
- Transfer Learning: To apply knowledge from well-mapped areas to underrepresented ones.
- Reinforcement Learning (Multi-armed Bandits): To balance "exploiting" known high-demand areas with "exploring" underserved ones to break feedback loops.
Key Insights: The Pothole Paradox and Autocorrelation
A primary example of these biases is the "Pothole Paradox." In 2012, a Boston app used smartphone accelerometers to detect potholes. Because smartphone ownership was higher in wealthy areas, the app "discovered" more potholes there, leading the city to fix roads in affluent neighborhoods while ignoring poorer ones. This is a classic "data glitch" where the sensor itself is biased.
Furthermore, the authors argue that spatial autocorrelation is a major hurdle for fairness metrics. Standard metrics often report "false positives" for bias because they don't account for natural human clustering. New metrics must be developed that distinguish between "natural" clustering and "unfair" discrimination.
Conclusion and Future Work
As a vision paper, the primary contribution is the strategic mapping of biases to mitigation strategies. The authors conclude that technical solutions must be paired with policy-making, using ML as a "guidance system" for policy makers to identify "broadband deserts" or "health deserts."
Future work involves developing specific "debiasing" layers for neural networks that can handle the 4D (3D space + 1D time) nature of this data without losing predictive accuracy. Key challenges remain, including the conflict between granular data collection for fairness and individual user privacy, as well as the need for political mechanisms to act on technical findings.
Summary updated