🤖 machine learning

Assessing Predictive Models for Fairness Based on Movement Patterns

This paper proposes a novel framework for assessing predictive model fairness by extending traditional spatial fairness concepts to incorporate individual movement patterns across multiple geographic regions, utilizing spatial scan statistics to effectively detect and localize movement-based biases.

Original authors: Francesco Lettich, Mario A. Nascimento, Chiara Pugliese, Chiara Renso

Published 2026-05-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Francesco Lettich, Mario A. Nascimento, Chiara Pugliese, Chiara Renso

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to figure out if a hiring manager (or a loan officer, or a police algorithm) is being unfair.

The Old Way: The "Home Address" Check
Traditionally, when people check for "spatial fairness," they look at where a person lives. They ask: "Does this algorithm treat people from Neighborhood A worse than people from Neighborhood B?"

It's like checking a guest list at a party. If the host only invites people from one specific street and ignores everyone else, that's unfair. But this method assumes everyone stays in one place, like a statue glued to their front porch.

The New Problem: The "Commuter" Reality
The authors of this paper point out that real people aren't statues. We move. We go to work, we stop at the grocery store, we visit the gym, and we commute through different neighborhoods.

Imagine a person who lives in a wealthy area but drives through a poorer neighborhood every day to get to work, stopping at a specific gas station and a small diner.

The Old Method might say: "This person lives in a rich area, so they are fine."
The Reality is: The algorithm might be looking at their daily route and the specific stops they make. If the algorithm sees that specific gas station and diner, it might unfairly label this person as "low income" or "high risk," even though they live in a rich area.

The paper argues that we need to check for fairness based on movement patterns, not just a single address. We need to ask: "Does the algorithm treat people unfairly because of the specific combination of places they visit?"

The Solution: The "Multi-Layered Map" Detective
To solve this, the authors built a new tool. Here is how it works, using a simple analogy:

The Stop-and-Go Filter: First, the tool looks at a person's movement history (like a GPS log) and ignores the times they are just driving fast. It only cares about the places they actually stop and stay for a while (like home, work, or the gym).
The Multi-Grid Net: This is the clever part. Imagine trying to find a lost coin in a park.
- If you look with a magnifying glass (a very fine grid), you see every single blade of grass, but it's hard to see the big picture.
- If you look with binoculars (a coarse grid), you see the whole park, but you might miss the coin hiding in the bushes.
- The authors' tool throws many nets of different sizes over the map at the same time. Some nets have tiny holes (fine detail), and some have huge holes (big picture). This ensures they don't miss unfairness just because the "unfair zone" doesn't fit perfectly into one specific map size.
The Pattern Match: The tool groups people based on the top 8 places they visit most often. It then asks: "Is there a specific group of people who visit this exact set of places, and does the algorithm treat them all differently than everyone else?"
The Statistical Alarm: If it finds a group that is being treated differently, it sounds an alarm. It doesn't just say "Unfairness exists"; it tries to pinpoint where on the map these unfair patterns are happening.

What They Found (The Experiment)
The authors couldn't test this on real people because they didn't know the "truth" (they didn't know for sure if the algorithm was actually biased). So, they created a video game world (a simulator) with 100,000 fake people living in Atlanta.

They programmed the "hiring manager" to be secretly unfair to specific groups of people who visited certain combinations of places. Then, they ran their new tool to see if it could catch the cheater.

Did it work? Yes. The tool was very good at sounding the alarm when unfairness was present.
Did it find the right people? Yes, it successfully identified the fake people who were being treated unfairly.
Did it find the exact location? This was tricky.
- Coarse maps (big grids) were great at finding that unfairness existed and finding the people involved.
- Fine maps (small grids) were better at drawing the exact boundaries of where the unfairness happened.
- The Strategy: The authors suggest a "zoom-in" approach. First, use the big nets to find the general area of trouble. Then, switch to the small nets to narrow down exactly where the problem lies.

The Bottom Line
This paper introduces a new way to check if algorithms are biased against people based on how they move through the world, not just where they sleep. It uses a "multi-resolution" strategy (looking at the map through different lenses) to ensure that no unfair pattern slips through the cracks, whether the unfairness is a tiny, specific spot or a large, sprawling area.

Technical Summary: Assessing Predictive Models for Fairness Based on Movement Patterns

Problem Definition
Current literature on spatial fairness in machine learning operates under a fundamental assumption: each individual is associated with a single, fixed geographical location (e.g., place of residence). This paper argues that this assumption is insufficient for real-world scenarios where individuals exhibit complex movement patterns. A predictive model may treat individuals unfairly not based on where they live, but based on their trajectories—specifically, the frequency, duration, and combination of regions they visit.

The authors define three dimensions of movement patterns that existing spatial fairness assessments fail to capture:

Multiple locations: Individuals occupy different regions at different times.
Frequency of visits: The number of times an individual visits specific regions matters.
Duration of visits: The length of time spent in a region is a relevant factor.

The core problem addressed is the assessment of predictive models for fairness relative to these movement patterns. Unlike traditional spatial fairness, where a "hotspot" of unfairness is a single contiguous region, unfairness based on movement patterns may manifest as a set of non-contiguous regions that an individual visits in a specific combination.

Methodology
The proposed approach generalizes spatial fairness assessment to handle movement data through a five-step pipeline:

Trajectory Segmentation: Trajectories are segmented into "stop" and "move" segments using a stop-and-move criterion. The approach focuses on stop segments, representing meaningful stays (e.g., home, work, shopping) rather than transient movements.
Superimposing Geographic Zonings: To mitigate the Modifiable Areal Unit Problem (MAUP), the method overlays multiple geographic zonings (spatial partitions) with varying resolutions and alignments over the area of interest.
Object-to-Cells Mapping: For each zoning, the approach maps moving objects to a subset of cells. It identifies stop centroids within cells and ranks them based on the number of distinct days a cell was visited. The top- $i$ cells (e.g., $i=8$ ) are selected to form a "cellset" representing the individual's relevant movement pattern.
Determining Subsets for Assessment: The method identifies all subsets of cells that appear in at least one individual's cellset. This is formulated as a Frequent Itemset Mining (FIM) problem, where cellsets are transactions and cells are items. This generates a candidate set of cell combinations ( $CANDIDATES_{GZ}$ ) to be tested for unfairness.
Hypothesis Testing: A spatial scan statistic (specifically, the Bernoulli-based statistic for binary classifiers in the experiments) is applied to the candidate subsets. The null hypothesis ( $H_0$ ) assumes the model is fair (global distribution of predictions is uniform). The alternative hypothesis ( $H_1$ ) posits that a subset of objects associated with a specific cell combination has a significantly different prediction distribution. The test uses Maximum Likelihood Ratio (MLR) and Monte Carlo simulations to determine statistical significance.

Key Contributions

Problem Formulation: The paper introduces the novel problem of assessing predictive models for fairness based on movement patterns, moving beyond the single-location assumption of spatial fairness.
Methodological Framework: It proposes a comprehensive approach that integrates trajectory segmentation, multi-resolution spatial zoning, frequent itemset mining for candidate generation, and spatial scan statistics for hypothesis testing.
Data Generation Protocol: The authors introduce a protocol for generating synthetic unfair auditable datasets. This involves injecting unfairness by manipulating labels for objects associated with specific movement patterns (hotspots composed of multiple regions), varying parameters such as hotspot shape, object count, and unfairness magnitude.
Exploratory Strategy: Based on experimental findings, the paper outlines a practical strategy for real-world assessments: pooling candidates across all resolutions to maximize detection power, then refining the search using coarse resolutions for object retrieval (sensitivity) and fine resolutions for precise spatial localization (Positive Predictive Value).

Experimental Results
The approach was evaluated on thousands of synthetic datasets containing injected unfairness. The evaluation focused on three metrics:

Statistical Power (Detection): The ability to detect the presence of unfairness.
Sensitivity: The ability to retrieve the set of objects treated unfairly.
Positive Predictive Value (PPV): The precision of the spatial localization of unfairness.

Key findings include:

Effectiveness: The approach effectively detects unfairness based on movement patterns and retrieves the affected objects.
Multi-Resolution Trade-off: There is a consistent trade-off between metrics and grid resolution. Coarser resolutions generally yield higher power and sensitivity (better at detecting and retrieving objects) but lower PPV (poorer spatial localization). Finer resolutions yield higher PPV (better localization) but may miss broader patterns.
Pooling Strategy: Pooling extreme candidates detected across all grid resolutions maximizes statistical power and sensitivity, serving as a robust initial step for detection.
Parameter Sensitivity: Detection performance improves with higher unfairness magnitude and larger numbers of objects per hotspot. However, increasing the complexity of hotspots (more regions or more stops) can complicate precise localization, though the approach remains effective at detection.

Significance and Claims
The authors claim that their work is the first to address the assessment of predictive models for fairness relative to individual movement patterns. They assert that existing spatial fairness methods may fail to detect unfairness if the "right" single location is not chosen, whereas their approach accounts for the aggregate movement behavior.

The paper modestly claims that the proposed approach is effective at detecting this new type of unfairness and retrieving the set of objects treated unfairly. It highlights that while localization performance exhibits trade-offs depending on spatial resolution, the proposed exploratory strategy offers a practical way to navigate these trade-offs in real-world assessments. The authors also note that their approach is generic and can reduce to standard spatial fairness assessment (Sacharidis et al., 2023) if individuals are tied to single fixed locations.

Limitations and Future Work
The authors acknowledge that their current implementation relies on uniform square grids and stop-segment-based association. Future research directions include investigating data-driven tessellations (e.g., DBSCAN), improving the reporting of candidate subsets for better localization, and exploring alternative criteria for associating objects with cells (e.g., considering move segments). Additionally, the protocol could be extended to other predictive models (regression, multi-class) and more varied movement data characteristics.

More like this