pith. machine review for the scientific record. sign in

arxiv: 2604.18644 · v2 · submitted 2026-04-19 · 💻 cs.LG · cs.AI

Recognition: unknown

FASE : A Fairness-Aware Spatiotemporal Event Graph Framework for Predictive Policing

Pronob Kumar Barman , Pronoy Kumar Barman , Plaban Kumar Barman , Rohan Mandar Salvi

Authors on Pith no claims yet

Pith reviewed 2026-05-10 06:01 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords predictive policingfairness constraintsspatiotemporal graphfeedback biaspatrol allocationcrime predictionclosed-loop simulation
0
0 comments X

The pith

Fairness constraints on patrol allocation alone leave a persistent detection gap in predictive policing simulations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper builds a framework that predicts crime in Baltimore neighborhoods using graph-based models and then allocates patrols under rules that keep demographic impact balanced. It runs repeated simulations in which the prediction model retrains on the crime reports that result from those patrols. Even though the allocation rules keep coverage high and demographic ratios near one, the rate at which crimes are detected stays about 3.5 percentage points lower in minority areas than in others. The authors conclude that fairness applied only at the allocation stage cannot break the feedback loop that produces biased training data.

Core claim

The FASE framework combines a spatiotemporal graph neural network and multivariate Hawkes process to forecast crime counts modeled as zero-inflated negative binomial distributions, then solves a linear program that maximizes risk-weighted patrol coverage subject to a Demographic Impact Ratio constraint with 0.05 deviation bound. Across six closed-loop deployment cycles on 2017-2019 Baltimore data, the ratio stays between 0.9928 and 1.0262 and coverage between 0.876 and 0.936, yet the detection-rate gap between minority and non-minority areas remains approximately 3.5 percentage points, showing that allocation-level constraints do not remove feedback-induced bias from retraining data.

What carries the argument

The closed-loop deployment feedback simulator that generates new training data from the outcomes of fairness-constrained patrol allocations and feeds it back into model retraining.

Load-bearing premise

The simulator's model of how patrol decisions and crime reporting produce feedback-driven changes in the data used for retraining matches real-world police and public behavior.

What would settle it

Running the fairness-constrained allocation rules in actual Baltimore policing operations and checking whether the 3.5-point detection-rate gap narrows after several retraining cycles.

Figures

Figures reproduced from arXiv: 2604.18644 by Plaban Kumar Barman, Pronob Kumar Barman, Pronoy Kumar Barman, Rohan Mandar Salvi.

Figure 1
Figure 1. Figure 1: fase pipeline overview. Arrows show data flow; the dashed feedback loop (Phase 5 → Phase 3) illustrates how ob￾served, patrol-biased counts are fed back into incremental model retraining. Feature-similarity adjacency Afeat: Cosine similarity of each ZCTA’s ACS feature vector (percentage minority, normalised me￾dian income, poverty rate), thresholded at 0.50, yielding 484 di￾rected edges (mean degree 19.36)… view at source ↗
Figure 3
Figure 3. Figure 3: Allocation DIR across 6 deployment cycles with the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Mean patrol units per ZCTA for minority vs. non [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 2
Figure 2. Figure 2: Phase 3 training dynamics over 100 epochs. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 7
Figure 7. Figure 7: ZINB retraining loss per cycle (mean and final [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Aggregate detection rates (observed / true counts) [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
read the original abstract

Predictive policing systems that allocate patrol resources based solely on predicted crime risk can unintentionally amplify racial disparities through feedback driven data bias. We present FASE, a Fairness Aware Spatiotemporal Event Graph framework, which integrates spatiotemporal crime prediction with fairness constrained patrol allocation and a closed loop deployment feedback simulator. We model Baltimore as a graph of 25 ZIP Code Tabulation Areas and use 139,982 Part 1 crime incidents from 2017 to 2019 at hourly resolution, producing a sparse feature tensor. The prediction module combines a spatiotemporal graph neural network with a multivariate Hawkes process to capture spatial dependencies and self exciting temporal dynamics. Outputs are modeled using a Zero Inflated Negative Binomial distribution, suitable for overdispersed and zero heavy crime counts. The model achieves a validation loss of 0.4800 and a test loss of 0.4857. Patrol allocation is formulated as a fairness constrained linear optimization problem that maximizes risk weighted coverage while enforcing a Demographic Impact Ratio constraint with deviation bounded by 0.05. Across six simulated deployment cycles, fairness remains within 0.9928 to 1.0262, and coverage ranges from 0.876 to 0.936. However, a persistent detection rate gap of approximately 3.5 percentage points remains between minority and non minority areas. This result shows that allocation level fairness constraints alone do not eliminate feedback induced bias in retraining data, highlighting the need for fairness interventions across the full pipeline.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. The paper claims to introduce the FASE framework for predictive policing, which uses a spatiotemporal graph neural network combined with a multivariate Hawkes process to model crime events in Baltimore based on 139,982 incidents from 2017-2019. It formulates patrol allocation as a linear optimization problem with a fairness constraint on the Demographic Impact Ratio (deviation ≤ 0.05), and employs a closed-loop simulator to run six deployment cycles. The results indicate that fairness is maintained between 0.9928 and 1.0262 with coverage from 0.876 to 0.936, but a 3.5 percentage point gap in detection rates between minority and non-minority areas persists, suggesting that allocation-level fairness constraints do not fully eliminate feedback-induced bias in retraining data.

Significance. If the closed-loop simulator accurately reflects real-world dynamics of patrol allocation influencing crime reporting and data bias, this work would be significant for demonstrating the insufficiency of isolated fairness interventions at the resource allocation stage in predictive policing. It provides a concrete example of how feedback loops can perpetuate disparities even under fairness constraints, advocating for comprehensive pipeline approaches. The integration of GNN and Hawkes process for spatiotemporal prediction is a technical contribution, but the overall impact is tempered by the lack of real-world validation.

major comments (3)
  1. [Simulation Results] The central claim that allocation-level fairness constraints alone do not eliminate feedback-induced bias rests on the observation of a persistent ~3.5pp detection-rate gap across six simulated deployment cycles. The manuscript provides no details on the simulator's mapping from patrol allocation to subsequent crime reporting rates, detection mechanics, or label generation for retraining data, nor any sensitivity analysis on parameters such as the 0.05 deviation bound or reporting elasticity.
  2. [Prediction Model] The prediction module reports a validation loss of 0.4800 and test loss of 0.4857 using a spatiotemporal GNN + Hawkes process with Zero-Inflated Negative Binomial outputs, but the manuscript supplies no information on data splits, hyperparameter search, or baseline comparisons. This information is required to establish that the risk predictions feeding into the fairness-constrained allocation are reliable and not themselves contributing to the observed gap.
  3. [Fairness Constraint Formulation] The Demographic Impact Ratio constraint (deviation bounded by 0.05) is presented as keeping fairness in [0.9928, 1.0262] while coverage ranges from 0.876 to 0.936, yet the manuscript does not specify how the 0.05 bound was chosen or demonstrate that the 3.5pp gap is robust to reasonable variations in this hyperparameter.
minor comments (3)
  1. The term 'Part 1 crime incidents' is used without definition or reference to its standard classification in the Baltimore dataset.
  2. Explicit mathematical notation for the sparse feature tensor, the Zero-Inflated Negative Binomial parameterization, and the Demographic Impact Ratio would improve clarity.
  3. A table summarizing all simulator parameters, their values, and justification would aid reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thoughtful and constructive comments. We agree that greater transparency on the simulator mechanics, prediction evaluation protocol, and fairness hyperparameter choices will strengthen the manuscript. We provide point-by-point responses below and will incorporate the requested clarifications and analyses in the revision.

read point-by-point responses

  1. Referee: The manuscript provides no details on the simulator's mapping from patrol allocation to subsequent crime reporting rates, detection mechanics, or label generation for retraining data, nor any sensitivity analysis on parameters such as the 0.05 deviation bound or reporting elasticity.

    Authors: We appreciate this observation. The current manuscript describes the closed-loop simulator at a high level but does not fully specify the functional mappings or parameter sensitivities. In the revised version we will add an expanded methods subsection that details the assumed relationships between patrol density and reporting/detection rates, the process for generating retraining labels from simulated detections, and a sensitivity analysis that varies the deviation bound and reporting elasticity parameters while tracking the persistence of the detection-rate gap. revision: yes

  2. Referee: The prediction module reports a validation loss of 0.4800 and test loss of 0.4857 using a spatiotemporal GNN + Hawkes process with Zero-Inflated Negative Binomial outputs, but the manuscript supplies no information on data splits, hyperparameter search, or baseline comparisons.

    Authors: We acknowledge the omission. The revised manuscript will include a dedicated evaluation subsection that describes the temporal data-splitting strategy employed to respect the time-series structure, the hyperparameter search procedure, and quantitative comparisons against relevant baselines (standard multivariate Hawkes process and non-graph spatiotemporal models). These additions will better substantiate that the risk predictions supplied to the allocation module are reliable. revision: yes

  3. Referee: The Demographic Impact Ratio constraint (deviation bounded by 0.05) is presented as keeping fairness in [0.9928, 1.0262] while coverage ranges from 0.876 to 0.936, yet the manuscript does not specify how the 0.05 bound was chosen or demonstrate that the 3.5pp gap is robust to reasonable variations in this hyperparameter.

    Authors: We selected the 0.05 bound to enforce a tight but operationally feasible fairness tolerance. The revision will explicitly state this rationale and add a supplementary experiment that re-runs the six-cycle simulation under alternative bounds (e.g., 0.03 and 0.07) to confirm that the observed detection-rate gap remains stable across reasonable choices of the hyperparameter. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected; results derive from independent simulation run

full rationale

The paper constructs a prediction model (STGNN + Hawkes + ZINB) on real Baltimore crime data, solves a constrained linear program for allocation, and then executes a separate closed-loop simulator to generate deployment cycles and measure downstream detection gaps. The 3.5pp gap is reported as an observed output of that simulator rather than being algebraically identical to any fitted parameter or input definition. No self-citations, uniqueness theorems, or ansatzes are invoked to force the central claim, and no equation reduces the fairness-bias conclusion to a tautology. The simulation is therefore an independent (if unvalidated) experimental apparatus, keeping the derivation self-contained.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The framework rests on standard assumptions about graph neural networks capturing spatial crime dependencies and Hawkes processes modeling self-excitation; the simulation itself introduces the feedback loop as an engineered construct rather than an independently validated mechanism.

free parameters (1)
  • Demographic Impact Ratio deviation bound = 0.05
    Explicitly set to 0.05 to enforce the fairness constraint in the linear program.
axioms (2)
  • domain assumption Spatiotemporal graph neural networks combined with multivariate Hawkes processes can adequately represent spatial dependencies and temporal self-excitation in crime event data.
    Invoked in the prediction module description.
  • domain assumption Zero-inflated negative binomial distribution is appropriate for modeling overdispersed, zero-heavy hourly crime counts.
    Stated as suitable for the data characteristics.

pith-pipeline@v0.9.0 · 5585 in / 1540 out tokens · 50725 ms · 2026-05-10T06:01:57.202936+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 16 canonical work pages

  1. [1]

    Kiana Alikhademi, Emma Drobina, Diandra Prioleau, Brianna Richardson, Dun- can Purves, and Juan E. Gilbert. 2022. A Review of Predictive Policing from the Perspective of Fairness.Artificial Intelligence and Law30, 1 (2022), 1–17. doi:10.1007/s10506-021-09286-4

    work page doi:10.1007/s10506-021-09286-4 2022

  2. [2]

    Almasoud and Jamiu Adekunle Idowu

    Ahmed S. Almasoud and Jamiu Adekunle Idowu. 2025. Algorithmic Fairness in Predictive Policing.AI & Ethics5, 3 (2025), 2323–2337. doi:10.1007/s43681-024- 00541-3

    work page doi:10.1007/s43681-024- 2025

  3. [3]

    Baltimore City Open Data. 2019. Baltimore Police Department Part 1 Crime Data, 2017–2019. https://data.baltimorecity.gov Accessed 2024

    2019

  4. [4]

    Pronob Kumar Barman and Pronoy Kumar Barman. 2025. Unmasking Algo- rithmic Bias in Predictive Policing: A GAN-Based Simulation Framework with Multi-City Temporal Analysis.arXiv preprint arXiv:2603.18987(2025). doi:10. 48550/arXiv.2603.18987

    work page arXiv 2025

  5. [5]

    Richard Berk, Hoda Heidari, Shahin Jabbari, Michael Kearns, and Aaron Roth

  6. [6]

    doi:10.1177/0049124118782533

    Fairness in Criminal Justice Risk Assessments: The State of the Art.Soci- ological Methods & Research50, 1 (2021), 3–44. doi:10.1177/0049124118782533

    work page doi:10.1177/0049124118782533 2021

  7. [7]

    Alexandra Chouldechova. 2017. Fair Prediction with Disparate Impact: A Study of Bias in Recidivism Prediction Instruments.Big Data5, 2 (2017), 153–163. doi:10.1089/big.2016.0047

    work page doi:10.1089/big.2016.0047 2017

  8. [8]

    Julia Dressel and Hany Farid. 2018. The Accuracy, Fairness, and Limits of Pre- dicting Recidivism.Science Advances4, 1 (2018), eaao5580. doi:10.1126/sciadv. aao5580

    work page doi:10.1126/sciadv 2018

  9. [9]

    Friedler, Scott Neville, Carlos Scheidegger, and Suresh Venkatasubramanian

    Danielle Ensign, Sorelle A. Friedler, Scott Neville, Carlos Scheidegger, and Suresh Venkatasubramanian. 2018. Runaway Feedback Loops in Predictive Polic- ing. InProceedings of the 1st Conference on Fairness, Accountability and Trans- parency (Proceedings of Machine Learning Research, Vol. 81). PMLR, 160–171. https://proceedings.mlr.press/v81/ensign18a.html

    2018

  10. [10]

    Andrew Guthrie Ferguson. 2017. Policing Predictive Policing.Washington Uni- versity Law Review94, 5 (2017), 1109–1189

    2017

  11. [11]

    Diego Hernández, Elena Santos, Xavier Ferrer, and Thomas Wischmeyer. 2026. Quantifying and Mitigating Racial Bias in Predictive Policing Algorithms.Arti- ficial Intelligence and Law(2026). doi:10.1007/s10506-026-09508-7

    work page doi:10.1007/s10506-026-09508-7 2026

  12. [12]

    Tzu-Wei Hung and Chun-Ping Yen. 2023. Predictive Policing and Algorithmic Fairness: An Analysis of Philosophical Foundations.Synthese201, 6 (2023), 206. doi:10.1007/s11229-023-04189-0

    work page doi:10.1007/s11229-023-04189-0 2023

  13. [13]

    Hyemi Kim, Seungjae Shin, Joon Ho Jang, Kyungwoo Song, Weonyoung Joo, Wanmo Kang, and Il-Chul Moon. 2021. Counterfactual Fairness with Disentan- gled Causal Effect Variational Autoencoder. InProceedings of the AAAI Confer- ence on Artificial Intelligence, Vol. 35. 8128–8136

    2021

  14. [14]

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. 2018. Diffusion Convolu- tional Recurrent Neural Network: Data-Driven Traffic Forecasting. InInterna- tional Conference on Learning Representations (ICLR). https://openreview.net/ forum?id=SJiHXGWAZ

    2018

  15. [15]

    Kristian Lum and William Isaac. 2016. To Predict and Serve?Significance13, 5 (2016), 14–19. doi:10.1111/j.1740-9713.2016.00960.x

    work page doi:10.1111/j.1740-9713.2016.00960.x 2016

  16. [16]

    Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. 2021. A Survey on Bias and Fairness in Machine Learning.Comput. Surveys54, 6 (2021), 1–35. doi:10.1145/3457607

    work page doi:10.1145/3457607 2021

  17. [17]

    Fox, Martin B

    George O. Mohler, Martin B. Short, Sean Malinowski, Mark Johnson, George E. Tita, Andrea L. Bertozzi, and P. Jeffrey Brantingham. 2015. Randomized Con- trolled Field Trials of Predictive Policing.J. Amer. Statist. Assoc.110, 512 (2015), 1399–1411. doi:10.1080/01621459.2015.1077710

    work page doi:10.1080/01621459.2015.1077710 2015

  18. [18]

    Schultz, and Kate Crawford

    Rashida Richardson, Jason M. Schultz, and Kate Crawford. 2019. Dirty Data, Bad Predictions: How Civil Rights Violations Impact Police Data, Predictive Policing Systems, and Justice.New York University Law Review Online94 (2019), 15–55. https://www.nyulawreview.org/online-features/dirty-data-bad-predictions/

    2019

  19. [19]

    Selbst, Danah Boyd, Sorelle A

    Andrew D. Selbst, danah boyd, Sorelle A. Friedler, Suresh Venkatasubramanian, and Janet Vertesi. 2019. Fairness and Abstraction in Sociotechnical Systems. InProceedings of the Conference on Fairness, Accountability, and Transparency (FAT*). ACM, 59–68. doi:10.1145/3287560.3287598

    work page doi:10.1145/3287560.3287598 2019

  20. [20]

    Samin Semsar, Kiran Laxmikant Prabhu, Gabriella Waters, and James Foulds

  21. [21]

    https://arxiv.org/abs/2602.02566

    A Comparative Simulation Study of the Fairness and Accuracy of Predic- tive Policing Systems in Baltimore City.arXiv preprint arXiv:2602.02566(2026). https://arxiv.org/abs/2602.02566

    work page arXiv 2026

  22. [22]

    Census Bureau

    U.S. Census Bureau. 2022. American Community Survey 5-Year Estimates, 2019. https://www.census.gov/programs-surveys/acs Accessed 2024

    2022

  23. [23]

    Caroline Wang, Bin Han, Bhrij Patel, and Cynthia Rudin. 2023. In Pursuit of Interpretable, Fair and Accurate Machine Learning for Criminal Recidivism Pre- diction.Journal of Quantitative Criminology39, 2 (2023), 519–581

    2023

  24. [24]

    Sheng Wang, Ying Zhang, Yang Liu, and Wei Chen. 2024. Fairness-Aware Po- lice Resource Allocation via Constrained Spatiotemporal Optimization.arXiv preprint arXiv:2403.00972(2024). https://arxiv.org/abs/2403.00972

    work page arXiv 2024

  25. [25]

    Tao Wang, Peng Chen, and Miaoxuan Shan. 2025. Crime Spatiotemporal Predic- tion Through Urban Region Representation by Using Building Footprints.Big Data and Cognitive Computing9, 12 (2025), 301

    2025

  26. [26]

    Jiahui Wu and Vanessa Frias-Martinez. 2024. Improving the Fairness of Deep- Learning Short-Term Crime Prediction with Under-Reporting-Aware Models. arXiv preprint arXiv:2406.04382(2024). https://arxiv.org/abs/2406.04382

    work page arXiv 2024

  27. [27]

    Zonghan Wu, Shirui Pan, Guodong Chen, Guodong Long, Chengqi Zhang, and Philip S. Yu. 2019. Graph WaveNet for Deep Spatial-Temporal Graph Modeling. InProceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI-19). 1907–1913. doi:10.24963/ijcai.2019/264

    work page doi:10.24963/ijcai.2019/264 2019

  28. [28]

    Junzhe Zhang and Elias Bareinboim. 2018. Fairness in Decision-Making—The Causal Explanation Formula. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 32

    2018

  29. [29]

    Marta Ziosi and Dasha Pruss. 2024. Evidence of What, for Whom? The Socially Contested Role of Algorithmic Bias in a Predictive Policing Tool. InProceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency. ACM, 1596–1608

    2024