Algorithmic fairness in predictive policing

Abstract

The increasing use of algorithms in predictive policing has raised concerns regarding the potential amplification of societal biases. This study adopts a two-phase approach, encompassing a systematic review and the mitigation of age-related biases in predictive policing. Our systematic review identifies a variety of fairness strategies in existing literature, such as domain knowledge, likelihood function penalties, counterfactual reasoning, and demographic segmentation, with a primary focus on racial biases. However, this review also highlights significant gaps in addressing biases related to other protected attributes, including age, gender, and socio-economic status. Additionally, it is observed that police actions are a major contributor to model discrimination in predictive policing. To address these gaps, our empirical study focuses on mitigating age-related biases within the Chicago Police Department's Strategic Subject List (SSL) dataset used in predicting the risk of being involved in a shooting incident, either as a victim or an offender. We introduce Conditional Score Recalibration (CSR), a novel bias mitigation technique, alongside the established Class Balancing method. CSR involves reassessing and adjusting risk scores for individuals initially assigned moderately high-risk scores, categorizing them as low risk if they meet three criteria: no prior arrests for violent offenses, no previous arrests for narcotic offenses, and no involvement in shooting incidents. Our fairness assessment, utilizing metrics like Equality of Opportunity Difference, Average Odds Difference, and Demographic Parity, demonstrates that this approach significantly improves model fairness without sacrificing accuracy.

Full text

Vol.:(0123456789)
AI and Ethics (2025) 5:2323–2337
https://doi.org/10.1007/s43681-024-00541-3
ORIGINAL RESEARCH
Algorithmic fairness inpredictive policing
AhmedS.Almasoud1· JamiuAdekunleIdowu2
Received: 31 January 2024 / Accepted: 29 July 2024 / Published online: 2 September 2024
© The Author(s) 2024
Abstract
The increasing use of algorithms in predictive policing has raised concerns regarding the potential amplification of societal
biases. This study adopts a two-phase approach, encompassing a systematic review and the mitigation of age-related biases
in predictive policing. Our systematic review identifies a variety of fairness strategies in existing literature, such as domain
knowledge, likelihood function penalties, counterfactual reasoning, and demographic segmentation, with a primary focus on
racial biases. However, this review also highlights significant gaps in addressing biases related to other protected attributes,
including age, gender, and socio-economic status. Additionally, it is observed that police actions are a major contributor to
model discrimination in predictive policing. To address these gaps, our empirical study focuses on mitigating age-related
biases within the Chicago Police Department's Strategic Subject List (SSL) dataset used in predicting the risk of being
involved in a shooting incident, either as a victim or an offender. We introduce Conditional Score Recalibration (CSR), a
novel bias mitigation technique, alongside the established Class Balancing method. CSR involves reassessing and adjust-
ing risk scores for individuals initially assigned moderately high-risk scores, categorizing them as low risk if they meet
three criteria: no prior arrests for violent offenses, no previous arrests for narcotic offenses, and no involvement in shooting
incidents. Our fairness assessment, utilizing metrics like Equality of Opportunity Difference, Average Odds Difference, and
Demographic Parity, demonstrates that this approach significantly improves model fairness without sacrificing accuracy.
Keywords Predictive policing· Algorithmic fairness· Mitigating age bias· Conditional score recalibration
1 Introduction
Over the past few years, the adoption of Artificial Intelli-
gence and Machine Learning technologies in the security
sector has seen a significant rise. A crucial application
within this area is predictive policing systems, which are
designed to predict potential criminal activities using data-
driven methodologies [23]. Aside from predicting where a
crime might take place, these algorithms can also identify
individuals who could potentially be involved in criminal
activities in the future. Furthermore, with the advancement
of computer vision in the security sector, AI applications
for real-time crime detection and prevention are increasing
[1, 29, 32]. Alongside predictive policing, recidivism algo-
rithms have also gained traction. These algorithms focus
on predicting the likelihood of a previously convicted indi-
vidual reoffending. Just like predictive policing, these algo-
rithms hold the promise of optimizing the justice system but
come with their own set of challenges especially bias which
needs to be addressed for AI to be a true enabler of sustain-
able development [14, 29].
The quality and representativeness of the data that
informs both predictive policing and recidivism algorithms
have become a point of contention. There are concerns about
data that may be skewed thereby leading algorithms to per-
petuate or even exacerbate these biases in their predictions
[6, 23]. This is evident in datasets used for crime detection,
which might not adequately represent all societal groups
[26]. Also, there have been instances in history when risk
assessments were influenced by sensitive features like race,
nationality, and skin color [4]. Beyond this, other factors like
* Jamiu Adekunle Idowu
ucabaid@ucl.ac.uk
Ahmed S. Almasoud
almasoud@psu.edu.sa
1 Department ofInformation System, Prince Sultan
University, Riyadh, SaudiArabia
2 Centre forArtificial Intelligence, Department ofComputer
Science, University College London, London,
UnitedKingdom
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2324 AI and Ethics (2025) 5:2323–2337
socio-economic status and age can also introduce unwanted
racial or demographic biases [28].
The implications of such biased algorithms are far-reach-
ing. They can result in marginalized communities facing
further disadvantages, or individuals suffering unwarranted
repercussions like false arrests based on AI's misclassifica-
tions [2, 9]. Recognizing these challenges, researchers have
proposed various mitigation strategies and fairness metrics.
Therefore, in our systematic review, we discussed the meth-
odologies proposed to enhance algorithmic fairness in both
predictive policing and recidivism. The primary objectives
are as follows:
1. To analyze the fairness metrics, sensitive features
and fairness strategies employed in existing literature
towards mitigating bias in policing algorithms.
2. To mitigate age-related biases in predictive policing
algorithms.
2 Objective 1: systematic literature review
In this section, we address our first research objective by
conducting a systematic literature review (SLR) to analyze
fairness metrics, sensitive features and fairness strategies
employed in existing literature. Our systematic literature
review includes papers that meet a set of predetermined eli-
gibility criteria.
2.1 Methods forSLR
We used the PRISMA guideline, updated in 2020 for the
review process [21]. The stages involved in this SLR are
outlined below:
2.1.1 Stage 1: data sources andsearch strategy
For the primary source of our systematic search, we chose
Scopus due to its vast collection of over 70 million publica-
tions, coupled with its high recall, precision, and reproduc-
ibility [34]. In addition, IEEE Xplore and ScienceDirect are
used as supplementary search systems [34]. The keywords
used for the search strategy are presented in Table1.
2.1.2 Stage 2: data sources andsearch strategy
This section sets out the criteria we applied when select-
ing studies for the systematic literature review. We screened
studies in a three-step process: initially examining titles and
keywords, followed by abstracts, and then a thorough read-
ing of the full text. Studies fitting our criteria proceeded to
Stage 3 for further evaluation.
2.1.3 Inclusion criteria
• Articles that directly investigate bias issues in policing
or recidivism algorithms.
• Publications from 2015 or later.
• Research materials including peer-reviewed journals,
conference papers, and white papers that have defined
research questions.
2.1.4 Exclusion criteria
• Non-research materials such as newsletters, magazines,
posters,invited talks, panels, keynotes, and tutorials.
• Duplicate research paper or article.
2.1.5 Stage 3: quality assessment
In our review, we used a specific set of questions to evaluate
the quality and relevance of each study. A total of five key
questions were employed to assess quality. Studies meeting
at least 3 out of these 5 questions were included for deeper
analysis.
QA1. Did the study have clearly defined research ques-
tions and methodology?
QA2. Did the study develop an algorithm specifically for
crime prediction or recidivism?
QA3. Did the study incorporate considerations of fairness
during the algorithm development process, ensuring non-
discrimination based on factors like gender, race, age,
etc.?
QA4. Was the study based on a real-world dataset?
Table 1 Search Keywords
Term Keywords
Fairness fair* OR bias OR trust* OR responsible OR equity OR "social justice" OR discrimination
AND
Police/recidivism police OR "law enforcement" OR policing OR "public safety" OR crim* OR recidivism
AND
Algorithm algorithm* OR "machine learning" OR models OR "artificial intelligence" OR predictive
OR automated
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2325AI and Ethics (2025) 5:2323–2337
QA5. Did the study report the performance of the algo-
rithm using suitable metrics such as accuracy, F1 score,
AUC-ROC, precision, recall, etc.?
2.2 Results
From the initial search, we identified 587 relevant stud-
ies. Out of these, 293 were sourced from Scopus, 185 from
SpringerLink, and 109 from IEEE Xplore. Using EndNote,
a reference management tool, we identified and removed
89 duplicate studies. This left us with 498 studies ready
for title and abstract screening. Upon examining the titles
and keywords, we excluded 426 studies. This led us to thor-
oughly read the abstracts of the remaining 72 articles. From
this group, 34 were found not to meet our criteria, leaving
38 for full-text examination. After this detailed review, 20
more articles were set aside, narrowing the list down to 18
as shown in Table2.
Following our quality assessment, only 15 of these stud-
ies were deemed suitable for our review. Figure1 provides
a visual representation of this selection process using a
PRISMA flow chart.
2.3 Discussion
2.3.1 Fairness metrics
This review uncovers a variety of methodologies and metrics
employed in the literature for measuring fairness of polic-
ing and recidivism algorithms, showcasing the complexity
Table 2 Papers selected for Quality Assessment Evaluation
Paper Reference Title
S1 [8] Evaluating Fairness in Predictive Policing Using Domain Knowledge
S2 [5] Improving fairness in criminal justice algorithmic risk assessments using optimal transport and conformal prediction sets
S3 [13] Predictive policing and algorithmic fairness
S4 [22] Data augmentation for fairness-aware machine learning: Preventing algorithmic bias in law enforcement systems
S5 [27] Is the data fair? An assessment of the data quality of algorithmic policing
S6 [20] Cohort bias in predictive risk assessments of future criminal justice system involvement
S7 [19] A penalized likelihood method for balancing accuracy and fairness in predictive policing
S8 [25] AI For Bias Detection: Investigating the Existence of Racial Bias in Police Killings
S9 [30] Accuracy and fairness in a conditional generative adversarial model of crime prediction
S10 [12] Artificial fairness? Trust in algorithmic police decision-making
S11 [31] Crowdsourcing perceptions of fair predictors for machine learning: A recidivism case study
S12 [16] Singular race models: addressing bias and accuracy in predicting prisoner recidivism
S13 [17] Algorithmic bias in recidivism prediction: A causal perspective
S14 [15] Accuracy, Fairness, and Interpretability of Machine Learning Criminal Recidivism Models
S15 [24] Case study: predictive fairness to reduce misdemeanor recidivism through social service interventions
S16 [3] A review of predictive policing from the perspective of fairness
S17 [7] Fairness, accountability and transparency: notes on algorithmic decision-making in criminal justice
S18 [33] Achieving equity with predictive policing algorithms: a social safety net perspective
Fig. 1 Flow diagram of study selection using PRISMA
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2326 AI and Ethics (2025) 5:2323–2337
and diversity of the field. For example, S1 employed five
distinct metrics; S2 used predictive parity; S9 implemented
the calibration test; S12 focused on False Positive Rates and
False Negative Rates, and S14 utilized fairness metrics from
the Aequitas toolkit. The five fairness metrics used by S1 are
briefly explained below:
• Statistical parity difference (SPD): Measures the differ-
ence in favorable outcomes between the unprivileged and
privileged groups. If SPD is zero, it indicates perfect fair-
ness, but a non-zero value indicates some level of dispar-
ity between the groups [8].
• Disparate impact (DI): Measures the ratio of favorable
outcomes for the unprivileged group to the favorable out-
comes for the privileged group. A value of 1 indicates
perfect fairness (i.e., both groups receive favorable out-
comes at the same rate). A value less than 1 suggests that
the unprivileged group is less likely to receive a favorable
outcome compared to the privileged group, while a value
greater than 1 indicates the opposite [8].
• Average odds difference: Measures the average difference
of false and true positive rate between the unprivileged
and privileged group [8].
• Equal opportunity difference: Measures the difference
of true positive rates between the privileged and unprivi-
leged groups [8].
• Theil index: A measure of the inequality in benefit alloca-
tion for individuals.
Meanwhile, S2 adopted prediction parity as it is good
for capturing group fairness [5]. On the other hand, S9
employed a calibration test to measure algorithmic fairness,
with residential income as the protected attribute. The fair-
ness measure involved assessing the calibration over a data
distribution, D. As an example, if 100 data points from D
predicted a crime occurrence with a 0.7 confidence level,
then the model's calibration is validated if 70 of its predic-
tions are accurate. This calibration assessment was executed
across ten bins of equal prediction accuracy [30].
Another eligible study, S12 utilized False Positive Rate
(FPR) and False Negative Rate (FNR) as metrics to identify
biases within their models. A high FPR implies a significant
number of non-recidivists being wrongly classified as poten-
tial reoffenders, leading to their undue detention. Conversely,
a high FNR indicates a considerable number of actual recidi-
vists incorrectly deemed fit for release, potentially resulting
in preventable crimes. The study also explored the implica-
tions of these rates for different subpopulations, highlighting
that biases manifest differently among groups. For instance,
they observed that African American group members often
exhibited characteristics of a higher false positive rate and a
lower false negative rate compared to other populations [16].
In their study, S14 leveraged several metrics from the
Aequitas toolkit. These metrics include Predicted Positive
Rate Disparity (PPRD), which measures whether the num-
bers of positive predictions are on par across groups. Also,
False Discovery Rate Disparity (FDRD) assessed the pro-
portion of false positives relative to all predicted positives,
while False Positive Rate Disparity (FPRD) focused on the
ratio of false positives to the actual negatives across groups.
Additionally, the study considered False Omission Rate
Disparity (FORD), evaluating the false negatives relative
to predicted negatives, and False Negative Rate Disparity
(FNRD), which assessed the ratio of false negatives com-
pared to the actual positives [15].
2.3.2 Datasets andsensitive features
In Table3, we make a comparison between the eligible
papers based on the datasets and sensitive features they
used. Most of the studies on Policing or Recidivism algo-
rithmic fairness are based on datasets from the United States.
They include Chicago’s Strategic List data (S1), New York
Police Department Crime Complaint data (S5), Indianapolis
Crime Incident data (S7), Police Killings data (S8), Criminal
Defendants of Broward County, Florida (S11), NIJ data from
Georgia, and Los Angeles’ Case Management data (S15).
The only exception is S9 which used crime incidents dataset
from SIEDCO, Bogotá, Colombia.
In terms of protected attribute being investigated, studies
gave major attention to race as 12 of the 15 studies compared
in the table investigated racial bias with the only exemp-
tion being S6, S9, and S10. Meanwhile, only two of the 15
studies investigated bias related to socio-economic status or
income (S9 and S11). Similarly, only four out of 15 papers
each examined bias related to gender (S1, S5, S11, and S14)
and three for age (S1, S5 and S6). In addition, only four of
the studies examined bias in more than one feature (S1, S5,
S11, S14).
2.3.3 Bias analysis andmitigation strategies
The analysis of the eligible papers shows that bias analysis
methods and mitigation strategies commonly employed in
policing and recidivism algorithms revolve around themes
like data-centric strategies (S1, S4, S5, & S8), counterfactual
reasoning (S2, S3, & S13), demographic segmentation (S6,
S12, & S15), and adding penalty to likelihood function (S7).
Table4 presents some of the bias analysis techniques and
mitigation strategies.
2.3.3.1 Data‑centric strategies (S1, S4, S5, andS8) Several
of the approaches directly target the dataset itself, reflecting
the maxim that "better data leads to better predictions."
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2327AI and Ethics (2025) 5:2323–2337
• Incorporation of relevant external data: S1 integrates
domain knowledge into a predictive policing dataset by
adding average rates of education, employment, and pov-
erty from the 2014 census, aiming to minimize bias and
fairness concerns. It found that integrating this domain
knowledge has minimal impact on classification metrics,
but notably improves fairness for all protected classes,
particularly in the Theil index for race [8].
• Data augmentation: To counteract bias in law enforce-
ment algorithms, S4 leverages a data augmentation strat-
egy for the RWF-2000 dataset, which showed an overrep-
resentation of dark-skinned males in violent situations.
Through techniques like Mask-RCNN and HRNet-w48,
the study synthetically oversamples undersampled enti-
ties to achieve balance, particularly focusing on race. The
results indicate that models trained on balanced datasets
generalized better, with increased accuracy and reduced
racial bias [22]. Similarly, S8 assigned class weights and
oversampled minority class to address data imbalance.
• Quality of the data: According to S5, data quality can
be as influential as the algorithm itself in determining
outcomes as missing or incomplete data can lead to
inaccurate or biased predictions. The paper's analysis of
NYPD’s Crime Complaint Data shows the data has a
high imbalance, which can affect the reliability of predic-
tions. Depending on the percentage of missing values,
it might be advisable to either conduct a complete case
analysis or discard certain variables from the dataset
[27]. For instance, according to the recommendations in
[18], if a dataset has less than 5% of its values missing, it
might be appropriate to conduct a complete case analysis.
This is because any resulting bias is likely, though not
guaranteed, to be minimal. On the other hand, if over
40% of the values for a particular variable are missing,
one might consider removing that variable from the anal-
ysis altogether.
2.3.3.2 Counterfactual reasoning and causal analysis (S2,
S3 andS13) S2 improved fairness in risk assessment algo-
rithms by employing counterfactual reasoning. First, it
trains a stochastic gradient boosting algorithm on a more
privileged protected class (White offenders). Then, it trans-
ports the joint predictor distribution from the less privileged
class (Black offenders) to the more privileged class with
the aim of treating the disadvantaged class members as if
they are members of an advantaged class [5]. By using opti-
mal transport, S3 removed the bias caused by differences
at arrangement between the joint predictor distributions for
Black and White offenders [13].
Table 3 Comparison of selected papers based on sensitive features
Paper Datasets Sensitive features
Sex Race Status* Age
S1 Strategic Subject List (SSL) dataset from Chicago Police Department listing arrest of 398,684 people from
August 1, 2012, to July 31, 2016
✓ ✓ ✓
S2 A random sample of 300,000 offenders in the United States ✓
S3 Strategic Subject List (SSL) dataset from Chicago Police ✓
S4 RWF-2000, which contains two thousand video clips lasting five seconds each, half of them depicting violent
acts and half of them normal activities
✓
S5 NYPD’s Crime Complaint Data (January–June 2018) published on NYC.gov ✓ ✓ ✓
S6 Data from Project on human development in Chicago covering 1,057 individuals from four different age
cohorts
✓
S7 Crime incident data from the city of Indianapolis, Indiana for the years 2012 and 2013. The four crime types
include aggravated assault, robbery, motor vehicle theft, and burglary
✓
S8 2013–2020 Police Killings” sheet from the Mapping Police Violence database ✓
S9 Dataset from SIEDCO that records all crime incidents in the city of Bogotá covering period 2016–2019. The
data set has 183,813 reports divided into thefts, injures, and homicides
✓
S10 Online experiment involving 642 UK residents recruited from an online crowdsourcing platform
S11 Public dataset containing information on criminal defendants of Broward County, Florida ✓ ✓ ✓
S12 Dataset assembled using the information and resources acquired from Florida Department of Corrections
(FDOC) and the Florida Department of Law Enforcement (FDLE)
✓
S13 Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) data offer 2years of data
(2013–2014)
✓
S14 NIJ contains information about 25,835 individuals from the state of Georgia who were granted parole from
2013–2015
✓ ✓
S15 Data extracts from the Los Angeles Attorney’s case management system were provided for the project ✓
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2328 AI and Ethics (2025) 5:2323–2337
Table 4 Bias Analysis and Mitigation Strategies
Paper Strategy/Analysis Implementation Result
S1 Incorporating domain knowledge Added features from external data (2014 census data)—aver-
age education rate, employment rate, poverty rate
Improved Theil index fairness for race while having minimal
effect on accuracy
S2 Counterfactual Reasoning for Improved Fairness Model trained on privileged class; Transfer of joint predictor
probability using optimal transport
Significant improvements in fairness for protected classes
S3 Causal Analysis for Model Discrimination Analysis of the causal relationships among various variables
impacting PPAs, particularly focusing on the role of police
actions
Police actions identified as the primary driver of model dis-
crimination in predictive policing
S4 Fairness-Aware Data Augmentation Identify biases; Oversample underrepresented entities in
video sequences
Improved accuracy; Reduced biases
S5 Data Quality Analysis Missingness analysis on NYPD crime complaint data Identified high missingness; highlighted importance of data
quality for fairness in algorithmic policing
S6 Cohort Bias Analysis Analyzed criminal histories across birth cohorts to assess
predictive risk assessments
Revealed overprediction of arrest risk for younger cohorts;
found bias across all racial group
S7 Penalized Likelihood Approach for Demographic Parity Added a penalty term to the likelihood function for police
patrols
Introduced fairness aligning patrol levels with demographics;
Accuracy decreased when prioritizing fairness
S8 Class weights and data oversampling Applied multiple ML models to an expansive police killing
dataset; used class weights and overspampling to correct
data imbalance
Achieved over 80% accuracy in classifying race of fatalities
S9 Conditional GANs Architecture for Crime Prediction ConvLSTM layers conditioned on past crime intensity maps,
weekdays, holidays
Improved spatiotemporal predictions; Reduced bias with mini-
mal accuracy loss
S10 Online experiment, text-based vignettes, random alloca-
tion to scenarios
Tested public trust in police decisions made by human vs.
algorithm, successful vs. unsuccessful outcomes, and indi-
vidual vs. area-based scenarios
People viewed decisions made by algorithms as less fair and
appropriate; successful algorithmic outcomes increased
support for general police use of algorithms; trust influenced
acceptability of algorithmic policing
S11 Crowdsourcing perceptions of fair predictors Recruited 90 crowdworkers to judge the inclusion of various
predictors for recidivism; used a bot to facilitate discussion
and voting
Diverse groups tended to agree more with the majority vote;
visual evidence was crucial in decision-making
S12 Singular Race Models Segment dataset based on race and Train race-specific models
with artificial neural networks
For most race-crime categories, improved accuracy over base
models; Magnitude of bias increased
S13 Causal reformulation of algorithmic fairness problem Used the Neyman-Rubin potential outcomes framework to
estimate algorithmic fairness from COMPAS data
Strong evidence of racial bias against African American
defendants
S14 Comparative analysis of ML models Created and evaluated various ML models (decision tree,
logistic regression, boosting classifiers, random forest,
SVM, neural network)
Found trade-offs between accuracy, fairness, and interpretabil-
ity; XGBoost had highest accuracy; Decision tree was most
interpretable and fairest for gender; Random forest was fairest
for both gender and race
S15 Post-hoc Bias Detection and Mitigation Adjusted score thresholds to balance recall across racial/eth-
nic groups in predictive models
Achieved better predictive fairness; identified trade-offs
between equity and efficiency
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2329AI and Ethics (2025) 5:2323–2337
Also, a causal analysis by S3 on predictive policing
algorithms (PPAs) found that police actions are the primary
factor contributing to model discrimination. The research
uncovered a vicious cycle: increased police deployment
leads to higher arrest rates, which then increases reported
crime rates, further justifying more police deployment [13].
The study considered a case study of the Chicago Police
Department with variables A (arrest records), O (PPA
output), I (PPA input and training), D (police actions), C
(reported crime rate), and relationships shown in Fig.2. A
vicious cycle emerges between D, A, and C (Fig.2), when
PPAs are not involved. When PPAs are integrated, complex
interactions among variables arise. Even if PPAs are banned,
effectively eliminating I and O, as recommended by [11], the
inherent bias in D persists. Therefore, D (police actions) is
the major driver of model discrimination, confirming [10]’s
findings that the racial disparity in Los Angeles Predictive
Policing Algorithms “lies in policing, not the algorithm.”
In addition, S13 analyzed the Correctional Offender Man-
agement Profiling for Alternative Sanctions (COMPAS) data
using the Neyman-Rubin potential outcomes framework for
causal inference. The analysis showed that the COMPAS
data has racial bias against African American defendants.
2.3.3.3 Demographic segmentation (S6, S12, andS15) S12
developed Singular Race Models, a novel approach to reduce
racially inspired bias in recidivism prediction, by segment-
ing the dataset based on race and training single race-based
models. This approach increased prediction accuracy and
analyzed race-related discrimination. However, it was noted
that bias still existed in the two race-based cohorts even in
the absence of race information, implying that many fea-
tures in the dataset are correlated with race [16]. Also, S15's
approach of developing individually tailored interventions
based on demography further emphasizes the importance
of understanding and catering to specific demographic
groups [24]. In addition, S6 highlighted an essential chal-
lenge known as cohort bias. By analyzing criminal histories
across different age cohorts, the study found that models
trained on older cohorts tend to over-predict the likelihood
of arrest for younger cohorts [20].
2.3.3.4 Fairness‑accuracy trade‑off (S4, S7, and S9) A
recurrent trend in these studies is the exploration of the bal-
ance between fairness and accuracy. S7 found a reduction
in algorithmic accuracy when demographic fairness was
achieved [19]. S9 introduced a fairness-accuracy balancing
technique, which quantified the tradeoffs between accuracy
and fairness of GANs model, successfully reducing bias
with a marginal impact on accuracy [30]. Similarly, the
data augmentation strategy by S4 increased accuracy and
reduced racial bias.
2.3.3.5 Others (S7, S9, 10, 11, and 14) S7 introduced a
penalized likelihood approach to achieve demographic par-
ity in crime models. The proposed strategy adjusts the like-
lihood function with a penalty term, ensuring police patrol
distribution among demographic groups is proportional to
their population representation. When applied to crime data
in Indianapolis, the model successfully implemented fair-
ness in patrol distribution, but this came with a minor reduc-
tion in prediction accuracy [19]. Meanwhile, another study,
S9 used a conditional GANs architecture to predict crimes
in Bogotá, Colombia, and identified potential bias affecting
vulnerable populations. To address fairness, S9 performed a
calibration test conditional on income level (high income vs
low income) as a protected variable [30]. Meanwhile, S10
examined public trust in algorithms versus human decision-
making in policing, finding that algorithmic decisions are
viewed as less fair and appropriate, yet successful use cases
can enhance overall acceptance. S11 focused on crowd-
sourcing perceptions of algorithmic fairness, with partici-
pants evaluating whether predictors are fair or not. S14 con-
ducted a comparative analysis of ML models for criminal
recidivism, highlighting trade-offs between accuracy, fair-
ness, and interpretability, with XGBoost showing the high-
est accuracy and decision trees being the most interpretable
and fairest in gender bias.
2.3.4 Human perceptions
One way to approach algorithmic bias in policing is to
include humans in the decision-making loop to bring a
layer of understanding, validation, and accountability to the
process. S10 investigated public perception and trust levels
towards algorithmic decision-making in policing, especially
when compared with decisions made by human officers. The
study conducted an online experiment involving 642 UK
residents recruited from an online crowdsourcing platform.
S10 found that people generally view decisions made by
algorithms as less fair and appropriate compared to those
made by police officers. However, this perception changes
Fig. 2 Casual relationships before (left) and after (right) PPA integra-
tion (S3)
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2330 AI and Ethics (2025) 5:2323–2337
with context. When exposed to instances of successful
algorithmic decision-making, trust in the algorithm and its
potential benefits increases. It is worth noting that the sam-
ple for this study was predominantly composed of younger
and White-British individuals, which could have potential
implications for the generalizability of these findings [12].
In a similar vein, S11 recruited 90 crowdworkers to iden-
tify predictors deemed 'fair' for algorithmic models. Par-
ticipants, in both individual and group settings, judged the
fairness of predictors for predicting recidivism. The study's
design, which utilized a popular workplace collaboration
tool, Slack, allowed for a structured yet dynamic exploration
of participant's opinions. The results highlighted that partici-
pants in a more structured group setting (with an appointed
leader) showed higher agreement with majority votes when
assessing predictor fairness. Additionally, there were some
predictors, like the defendant's race, that were identified for
exclusion in both participants didn't seem to influence their
voting behavior on demographic predictors, suggesting that
individual biases may not have played a significant role in
the decision-making process [31].
2.4 Recommendations based onSLR
In our systematic literature review, we have investigated fair-
ness in policing and recidivism algorithms. The analysis of
the selected papers revealed a range of strategies aimed at
addressing inherent biases in these algorithms, highlighting
the multidimensional nature of fairness in the domain. Based
on our analysis, we make the following recommendations for
researchers to consider in future:
1. Human validation in algorithmic processes: Considering
findings from S10 and S11, it is recommended to incor-
porate human judgment in algorithmic processes. This
ensures decisions benefit from both human intuition and
machine efficiency. Additionally, structured group set-
tings, as observed in S11, can be adopted for consistent
fairness evaluations.
2. Expand datasets beyond the US: Given the predomi-
nance of U.S-based datasets as observed in studies like
S1, S2, S3, S5, S6, S7, S11, S12, S14, and S15, research-
ers are encouraged to diversify geographical sources to
enhance global relevance as the societal norms and legal
structures influencing criminal behavior vary widely
across countries [37]. Also, the U.S. context often dif-
fers from other regions in terms of legal systems, socio-
economic conditions, and demographic compositions.
Therefore, relying solely on U.S-based datasets can limit
the generalizability of findings and fairness solutions.
3. Expand the scope of protected attributes in bias investi-
gations: The emphasis on racial bias in the majority of
the studies, as observed from 12 out of 13 papers refer-
enced, highlights a narrow focus in bias investigations.
In contrast, other critical attributes like socio-economic
status or income, gender, and age received significantly
less attention, with only two studies examining bias
related to socio-economic status (S9 & S11) and only
three studies each for gender (S1, S5 & S11) and age
(S1, S5 & S6). Therefore, to ensure fairness across mul-
tiple dimensions with better generalizability of algorith-
mic solutions, it's essential for future research to pay
attention to gender, socio-economic status, and age in
addition to race.
4. Tailor bias mitigation based on context: Given the
diverse strategies highlighted in S1, S2, S4, S7, S9, S12,
and S18, it's advisable for researchers to critically ana-
lyze biases based on the specific context and application
of policing and recidivism algorithms. Bias mitigation
should also be tailored to fit the unique context, ensuring
more relevant and effective fairness outcomes.
3 Objective 2: mitigating age‑related bias
In this section, we implement the second objective of our
paper which involves investigating and mitigating bias
related to age in policing algorithms. The policing algorithm
developed in this study is for predicting the risk of involve-
ment in a shooting incident either as a victim or an offender
and it is based on the Chicago Police Department’s Strategic
Subject List.
As highlighted in our SLR, age, socio-economic status,
and gender are sensitive features often overlooked by previ-
ous studies on predictive policing. For instance, only 3 out
of the 15 papers in our SLR examined bias related to age, 4
examined gender bias, and only 2 investigated bias related
to socio-economic status compared to 12 out of 15 for race.
While any of these understudied categories could be chosen
towards addressing the gap identified in our SLR, the choice
of focusing on age-related bias in this study is driven by the
characteristics of the dataset being analyzed. Specifically,
the Chicago Police Department’s Strategic Subject List con-
tains age and gender but lacks socio-economic status data.
Additionally, the dataset shows significant bias with respect
to age as reported in [8] and [37]. Therefore, we are focusing
on mitigating age-related bias in this study. However, this
does not negate the importance of socio-economic status
or gender; rather, it addresses one part of the gaps while
recognizing that others remain and should be explored in
future research.
3.1 Dataset
This research used the Chicago Police Department’s Strate-
gic Subject List dataset which covers 398,684 people from
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2331AI and Ethics (2025) 5:2323–2337
August 2012 to July 2016. The Strategic Subject List (SSL)
score is a risk assessment score that reflects an individual’s
probability of being involved in a shooting incident either as
a victim or an offender. The scores range from 0 (extremely
low risk) to 500 (extremely high risk). According to [36],
scores above 250 are deemed to be high risk. According to
the Chicago Police Department, the SSL is intended to "rank
individuals according to their probability of being involved
in a shooting incident, either as an offender or a victim,"
with scores recalculated daily.
Meawhile, there is a huge discrepancy in the dataset that
points to age bias as analysis revealed that age accounts for
about 89% of the variance in SSL scores [36]. Specifically,
127,513 individuals have never been arrested or shot, but
around 90,000 of them are deemed to be at high risk.
3.1.1 Data pre‑processing
Initially, the dataset has 398,684 data instances and 48 fea-
tures. We applied preliminary pre-processing steps which
involved dropping columns with significant missing values,
dropping rows with missing values for longitude and latitude
columns, among others. Afterwards, we followed up with
additional pre-processing steps highlighted below:
1. The feature 'AGE AT LATEST ARREST' originally con-
tained age ranges. We simplified this by categorizing
individuals into two groups: those under 30years old
('less than 20' and '20–30') were encoded as 1 and those
30years or older were encoded as 0. This binary clas-
sification is based on our initial findings that the most
pronounced differences in SSL scores were between
these two age groupings. One critical consideration is
that the split needs to be relatively balanced. The data
points for each age range before the binary split were as
follows:
• Less than 20 yrs: 41,415 individuals
• 20–30 yrs: 76,731 individuals
• 30–40 yrs: 46,915 individuals
• 40–50 yrs: 32,587 individuals
• 50–60 yrs: 21,198 individuals
• 60–70 yrs: 4,564 individuals
• 70–80 yrs: 500 individuals
Combining 'less than 20' and '20-30' (to make <30yrs)
results in a total of 118,146 individuals, while the older
age groups (>30yrs) sum to 105,764, ensuring a rela-
tively balanced split.
2. For the 'RACE CODE CD' feature, we retained only two
categories for simplicity and relevance to the study's
focus: 'BLK' (Black) and 'WHI' (White), encoded as 1
and 0, respectively. Rows of data that did not fall into
these two categories were dropped.
3. The 'WEAPON I' and 'DRUG I' features were binary
encoded 1 (True) and 0 (False).
4. For the target variable, SSL SCORE, scores above 250
were encoded as 1 indicating high risk, while scores of
250 and below were encoded as 0, indicating low risk
aligning with [36].
In the end, we have 170,694 data instances and 12 fea-
tures. The features are listed in Table5.
Additional feature engineering: After getting initial
results for our models, we obtained feature importance and
based on the values for each feature, we decided to drop
WEAPON I, DRUG I, LATITUDE I, AND LONGITUDE
I from the features as they have very little contribution to
the model.
Table 5 Features selected as predictors
Feature Description
Age at latest arrest The individual's age at the time of their most recent arrest
Victim shooting incidents The number of times an individual has been the victim of a shooting
Victim battery or assault The number of times an individual has been the victim of an aggravated battery or
aggravated assault;
Arrests violent offenses The number of times the individual has been arrested for a violent offense
Gang affiliation Indicator if an individual has been confirmed to be a member of a criminal street gang
Narcotic arrests The number of times the individual has been arrested for a narcotics offense
Trend in criminal activity The trend of an individual's recent criminal activity
UUW Arrests The number of times the individual has been arrested for Unlawful Use of Weapons
Weapon I Is ‘Yes’ if at least one Weapon (UUW) Arrest in past 10years
Drug I Is ‘Yes’ if at least one Drug Arrest in past 10years
Latitude Latitude of Centroid of Census Tract of Arrest for the Subject's latest arrest record
Longitude Longitude of Centroid of Census Tract of Arrest for the Subject's latest arrest record
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2332 AI and Ethics (2025) 5:2323–2337
Sensitive Features: The sensitive features in this research
are AGE and RACE. However, it should be noted that RACE
is only used for bias assessment. It is not included as a pre-
dictor in this task just like the dataset provider (Chicago
Police Department) excluded it.
3.2 Model selection andevaluation
3.2.1 Model selection
The models developed in our study serve a critical role in
predicting the risk of being involved in a shooting incident,
either as a victim or an offender. Specifically, the models
predict a high risk (1) if the predicted risk score of the sub-
ject is more than 250, and a low risk (0) if the risk score
is less than 250. Accurately predicting the risk of involve-
ment in shooting incidents helps law enforcement agencies
in prioritizing surveillance, intervention efforts, and resource
allocation, thereby improving public safety. By mitigating
age-related bias in the algorithms, we ensure that individuals
are not unfairly targeted based on biased assessments.
The models selected for this task are:
• Random forest: Random Forest is an ensemble learning
method that constructs multiple decision trees during
training and outputs the mode of the classes (classifica-
tion) of the individual trees. It is known for its robust-
ness against overfitting and ability to handle non-linear
relationships.
• Logistic regression: Logistic Regression is a linear model
used for binary classification tasks. It models the prob-
ability that a given input point belongs to a certain class.
Despite its simplicity, logistic regression is highly inter-
pretable and can serve as a strong baseline model.
• Gradient boosting: Gradient Boosting is an ensemble
technique that builds models sequentially, with each new
model attempting to correct the errors of the previous
ones. This approach typically results in models that are
more accurate and less prone to overfitting than indi-
vidual models.
3.2.2 Performance metrics
To evaluate the performance of the models, we used the
following metrics:
• Accuracy: Accuracy is the proportion of correctly pre-
dicted instances (both true positives and true negatives)
out of the total instances. It provides a straightforward
measure of the models' overall performance.
• F1 Score: The F1 score is the harmonic mean of preci-
sion and recall. It considers both false positives and false
negatives, providing a balanced measure of the models'
accuracy, particularly useful for imbalanced datasets.
3.2.3 Fairness metrics
The fairness metrics used for evaluation of the models
include Equality of Opportunity Difference, Average Odds
Difference and Demographic Parity. For all the fairness met-
rics, the prediction is considered fair is the value falls within
the acceptable threshold of -0.1 to 0.1.
1. Demographic parity: This metric measures whether the
probability of a positive outcome (e.g., being classified
as high risk) is the same across different groups. In other
words, demographic parity is achieved if each group
has an equal chance of receiving the positive outcome,
regardless of their actual proportion in the population.
2. Equality of opportunity: This metric specifically focuses
on the true positive rate, ensuring that all groups have
an equal chance of being correctly identified for the
positive outcome when they qualify for it. It is a more
nuanced metric that considers the accuracy of the posi-
tive predictions for each group, thereby ensuring fairness
in the model's sensitivity.
3. Average odds difference: This metric evaluates the aver-
age difference in the false positive rates and true positive
rates between groups. It combines aspects of both false
positives (instances wrongly classified as positive) and
true positives (correctly classified positive instances),
providing a comprehensive view of the algorithm's per-
formance across different groups. A lower value in this
metric indicates a fairer algorithm, as it suggests mini-
mal disparity in both false and true positive rates across
groups.
3.3 Bias analysis andmitigation
According to [8], the Strategic Subject List dataset is highly
skewed and has bias with respect to age. People over the
age of 40 tend to have a low SSL score. In addition, while
it is unlikely that all the people under 20years old would
be involved in a shooting, the dataset showed all of them as
high risk [8]. Furthermore, there are 127,513 individuals
on the list who have never been arrested or shot, but around
90,000 of them are deemed to be at high risk [36]. Therefore,
it’s critical to analyze the bias in the SSL dataset before
building the algorithms.
1. The core of our analysis involves assessing the distribu-
tion of SSL scores across the demographic categories
i.e., age, gender, and race. We did this analysis using
histogram plots as shown in Figs.3, 4, and 5.
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2333AI and Ethics (2025) 5:2323–2337
2. Second, we examined a specific subset of dataset cap-
turing individuals with SSL scores over 250 who are
classified as “High Risk” despite having no prior arrest
for violent offenses, no prior arrest for narcotic offenses
and never been a victim of shooting incidents.
3.3.1 Conditional score recalibration (CSR)
Our preliminary analysis of the dataset shows that all indi-
viduals below the age of 30 have scores higher than 250,
that is, they are all marked as High Risk. This shows that the
dataset is significantly skewed against young people. How-
ever, rather than implement a mitigation strategy that will
Fig. 3 Distribution of SSL
Scores by Race
Fig. 4 Distribution of SSL
Scores by Age
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2334 AI and Ethics (2025) 5:2323–2337
focus on young people alone (as that would likely introduce
bias against old people), we introduced a more nuanced miti-
gation strategy, Conditional Score Recalibration which was
applied to the entire dataset, young and old. This strategy
is designed to reassess and adjust the risk scores assigned
to individuals within a specific subset of the dataset. Under
the Conditional Score Recalibration framework, individuals
originally assigned a moderately high-risk score, ranging
from 250 to 350, are re-evaluated and marked as low risk if
they meet ALL the following conditions:
a) No prior arrest for violent offenses
b) No prior arrest for narcotic offenses
c) Never been a victim of shooting incidents.
3.3.2 Class balancing
Post-recalibration, the dataset comprised 111,117 individu-
als classified as low risk and 59,577 individuals classified
as high risk. To address this imbalance, we employed a
class balancing technique. This technique involved under-
sampling the majority class, which in this case was the low-
risk group. By randomly selecting a subset of the low-risk
individuals equivalent in size to the high-risk group, we
achieved a balanced dataset.
3.4 Results
Figure3 illustrates a clear racial disparity within the dataset,
with a significantly higher number of black individuals com-
pared to white. This discrepancy suggests that Black indi-
viduals are more frequently involved in arrests or encounters
with the police.
Figure4 reveals a notable difference in the age demo-
graphic within the dataset compared to the trends seen in
race. First, the dataset comprises a higher proportion of
younger individuals (under 30years) compared to older indi-
viduals (30years and above). Furthermore, a striking aspect
revealed by the analysis is that all individuals in the younger
age bracket are assigned SSL scores above 250, categorizing
them as high risk. In contrast, the older age group shows a
markedly different distribution as only a small fraction of
individuals in this demographic are classified as high risk,
with the majority being deemed low risk.
This disparity in risk classification between younger and
older individuals raises critical considerations about the
underlying factors driving these assessments. The 100%
classification of younger individuals as high risk, regard-
less of other factors, may point to age-related biases in the
Chicago Police department assessment process. This finding
aligns with [36] which stated that the SSL scores assigned
by the Chicago Police Department do little more than rein-
force the “age out of crime” theory – a theory that suggests
individuals “grow out” of crime in their 30s. No doubt,
this shows that the dataset is highly biased against young
people. It also aligns with findings by [13] which noted that
police actions constitute the primary contributor to model
discrimination.
Figure5 from our analysis further highlights a significant
aspect of the dataset. It shows a considerable number of
individuals who have not had any prior arrests for violent or
narcotic offenses and have never been victims of shooting
incidents, yet they have been assigned SSL scores above
250, categorizing them as high risk. This finding raises
important questions about the criteria used for determining
SSL scores and suggests a potential disconnect between an
Fig. 5 SSL Scores (> 250) for
those with no violent & narcotic
arrests and shooting incidents
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2335AI and Ethics (2025) 5:2323–2337
individual's documented criminal history and their assessed
risk level.
To determine the most suitable model for our analysis,
we initially developed three different models: Logistic
Regression, Random Forest, and Gradient Boosting, with-
out incorporating any bias mitigation strategies. The per-
formance of all three models was comparable; however,
we ultimately selected the Random Forest model due to its
robustness against overfitting and its ability to handle non-
linear relationships in large datasets more effectively. The
performance and fairness result of the Random Forest model
are in Table6.
These results reveal that the model performs fairly with
respect to race, as all fairness metrics for race are within the
generally accepted threshold of −0.1 to 0.1. However, there
is a significant bias regarding age indicating a pronounced
disparity in the model's predictions. This high model bias
against young people is not a surprise given the distribution
of SSL Scores by Age in Fig.4. After implementing the
Conditional Score Recalibration and Class balancing strate-
gies to mitigate the bias, we obtained the following results
(Table7):
The post-mitigation results show a significant improve-
ment in the model's fairness with respect to age. The Demo-
graphic Parity metric for age decreased significantly from
0.8517 to 0.3128, and the Equality of Opportunity metric
saw a reduction from 0.7616 to 0.1521. The Average Odds
Difference for age also improved from 0.3349 to a much
lower value of 0.02024, indicating a substantial reduction
in bias. For race, the results remained largely within the
acceptable threshold, with the Equality of Opportunity and
Average Odds Difference metrics well within the −0.1 to
0.1 range. The Demographic Parity metric, while slightly
exceeding this range at 0.117, is still relatively close, indicat-
ing that the model maintains a fair degree of parity across
racial groups.
3.5 Discussion
Using biased datasets without mitigation strategies can per-
petuate systemic biases, leading to discriminatory practices
and unjust outcomes for affected individuals. In the context
of predictive policing, such biases can undermine public
trust in law enforcement and worsen existing inequalities
[35]. By mitigating these biases, we not only promote fairer
treatment of individuals but also enhance the legitimacy and
effectiveness of law enforcement practices.
In our study, we observed significant age-related biases
in the Chicago Police Department's SSL dataset. From the
perspective of theories of justice, particularly John Rawls'
theory of justice as fairness, this unequal treatment of dif-
ferent age groups is problematic. Rawls' theory advocates
for the arrangement of social institutions to benefit the least
advantaged members of society [40]. In this context, having
all 118,146 individuals under the age of 30 categorized as
high risk in our dataset represents an unfair treatment. The
ethical imperative, therefore, is to design and implement
algorithms that do not inherently disadvantage any specific
demographic group based on protected attributes like age.
The CSR introduced in our study and the well-established
Class Balancing technique provide a framework for achiev-
ing this goal.
The implementation of bias mitigation strategies not
only improved fairness but also improved the model's over-
all performance. The accuracy of the model increased from
0.8314 to 0.9014, and the F1 score increased from 0.83 to
0.90. This improvement effectively dispels the commonly
held notion that efforts to increase fairness invariably com-
promise performance. In fact, our findings are in line with
recent research on policing and recidivism algorithms, such
as [22] and [30], which report that there is no strict trade-off
between fairness and accuracy.
Aside from fairness and accuracy, other pillars of respon-
sible AI include privacy, interpretability or explainability,
and transparency. Data collection, such as the Chicago
Police Department monitoring individuals and collecting
their data without consent, raises privacy concerns. How-
ever, another school of thought may argue that such surveil-
lance is for the greater good as a secure society outweighs
personal privacy concerns. Also, relying solely on model
decisions is insufficient as findings by [12] and [31] show it
is critical to incorporate human validation, where humans
review the top features driving the model's recommenda-
tions and make the final decisions. Yet, this introduces
another layer of complexity, as humans are inherently biased
and may introduce additional biases. Thus, it is essential
Table 6 Results obtained for Random Forest Model
Accuracy = 0.8314, F1 = 0.83
Sensitive
features
Demographic parity Equality of
opportunity
Average
odds differ-
ence
Race 0.09923 0.08356 0.07679
Age 0.8517 0.7616 0.3349
Table 7 Results obtained for Random Forest after applying bias miti-
gation strategies
Accuracy = 0.9014, F1 = 0.90
Sensitive
features
Demographic
parity
Equality of oppor-
tunity
Average
odds differ-
ence
Race 0.1170 0.08768 0.03523
Age 0.3128 0.1521 0.02024
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2336 AI and Ethics (2025) 5:2323–2337
to analyze the trade-offs between these pillars critically.
Increasing transparency and explainability, for example,
can help uncover the black-box nature of models, thereby
enhancing trust among officers and the public [39]. How-
ever, this level of transparency might require compromising
some degree of data privacy.
While our study mitigated age-related bias, it has some
limitations as presented below:
1. Our SLR revealed that most existing studies on predic-
tive policing used US datasets. In deed, the Chicago
Police Department's SSL dataset reflects the specific
socio-economic, cultural, and policing practices of
a single city in the United States. This geographical
limitation means that our findings may not be directly
applicable to other regions and countries with differ-
ent policing strategies, demographic compositions, and
social contexts. Future research should incorporate data-
sets from various geographical locations to validate the
findings and extend generalizability.
2. The binary classification of age groups into ‘ < 30’ and
‘ > 30’ simplifies the age attribute but may overlook
nuances within these groups. It will be critical for future
research to consider the granular details and analyze
each age range to uncover additional insights.
3. Our SLR reveals that very few studies focus on biases
related to socio-economic status, age, and gender com-
pared to race. While this study has focused on mitigating
age-related bias, it is critical for future studies to con-
sider socio-economic status and gender biases as well.
4. The Conditional Score Recalibration (CSR) introduced
in our study has shown promising results in reducing
age-related biases. However, CSR relies on predefined
criteria for recalibrating risk scores, which might not
capture all dimensions of an individual's risk profile.
4 Conclusion
This study involves a systematic literature review on the
fairness of predictive policing algorithms and mitigating
age-related bias in predictive policing. From our SLR, we
found out the need to expand the scope of protected attrib-
utes beyond commonly researched area like race to include
gender, age, and socio-economic status. Responding to this
and leveraging the Chicago Police Department's Strategic
Subject List (SSL) dataset in predicting the risk of inovlve-
ment in a shooting incident, we analyzed and mitigated age-
related bias in the SSL dataset, particularly against younger
individuals. Our approach involves two strategies namely
Conditional Score Recalibration and Class Balancing.
We observed that the application of these strategies led
to a significant reduction in age-related biases. Remarkably,
this improvement in fairness did not come at the cost of
accuracy. Instead, we witnessed an increase in the model's
accuracy from 0.8314 to 0.9014, and an improvement in the
F1 score from 0.83 to 0.90. These results challenge the preva-
lent notion that fairness and accuracy are mutually exclusive.
Also, the study contributes to the ongoing discourse on the
responsible use of AI in law enforcement, emphasizing the
importance of continuously scrutinizing and refining predic-
tive policing tools to ensure they serve the public equitably.
Acknowledgement The authors acknowledge the support of University
College London and Prince Sultan University.
Data availability The dataset used for this study can be found at https://
data. cityo fchic ago. org/ Public- Safety/ Strat egic- Subje ct- List- Histo rical/
4aki- r3np
Declarations
Conflict of interest The authors declare no conflicts of interest relevant
to this work.
Open Access This article is licensed under a Creative Commons Attri-
bution 4.0 International License, which permits use, sharing, adapta-
tion, distribution and reproduction in any medium or format, as long
as you give appropriate credit to the original author(s) and the source,
provide a link to the Creative Commons licence, and indicate if changes
were made. The images or other third party material in this article are
included in the article’s Creative Commons licence, unless indicated
otherwise in a credit line to the material. If material is not included in
the article’s Creative Commons licence and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder. To view a
copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
References
1. AbdElaziz, M., Al-qaness, M.A., Dahou, A., Ibrahim, R.A., Abd
El-Latif, A.A.: Intrusion detection approach for cloud and IoT
environments using deep learning and capuchin search algorithm.
Adv. Eng. Softw.Softw. 176, 103402 (2023). https:// doi. org/ 10.
1016/j. adven gsoft. 2022. 103402
2. Abdelkader, M., Mabrok, M., Koubaa, A.: OCTUNE: optimal
control tuning using real-time data with algorithm and experi-
mental results. Sensors 22(23), 9240 (2022). https:// doi. org/ 10.
3390/ s2223 9240
3. Alikhademi, K., Drobina, E., Prioleau, D., Richardson, B., Purves,
D., Gilbert, J.E.: A review of predictive policing from the perspec-
tive of fairness. Artif. Intell. Law. Intell. Law. (2022). https:// doi.
org/ 10. 1007/ s10506- 021- 09286-4
4. Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016). Machine
bias, ProPublica. Retrieved April 19, 2019.
5. Berk, R.A., Kuchibhotla, A.K., TchetgenTchetgen, E.: Improving
fairness in criminal justice algorithmic risk assessments using
optimal transport and conformal prediction sets. Sociol. Method.
Res. (2021). https:// doi. org/ 10. 1177/ 00491 24123 1155
6. Buolamwini, J., & Gebru, T. (2018, January). Gender shades:
Intersectional accuracy disparities in commercial gender classi-
fication. In Conference on fairness, accountability and transpar-
ency (pp. 77–91). PMLR.
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

2337AI and Ethics (2025) 5:2323–2337
7. Chiao, V.: Fairness, accountability and transparency: notes on
algorithmic decision-making in criminal justice. Internat. J. Law
Context 15(2), 126–139 (2019)
8. Downey, A., Islam, S. R., Sarker, M. K. (2023). Evaluating Fair-
ness in Predictive Policing Using Domain Knowledge. In The
International FLAIRS Conference Proceedings. https:// doi. org/
10. 32473/ flairs. 36. 133088
9. Fang, S., Chen, H., Khan, Z., Fan, P.: User fairness aware power
allocation for NOMA-assisted video transmission with adaptive
quality adjustment. IEEE Trans. Veh. Technol.Veh. Technol. 71(1),
1054–1059 (2021). https:// doi. org/ 10. 1109/ TVT. 2021. 31298 05
10. Ferguson, A.G.: Policing predictive policing. Wash. UL Rev. 94,
1109 (2016)
11. Heaven, W. D. (2020). Predictive policing algorithms are racist.
They need to be dismantled. MIT Technology Review, 17, 2020.
12. Hobson, Z., Yesberg, J.A., Bradford, B., Jackson, J.: Artificial
fairness? Trust in algorithmic police decision-making. J. Exp.
Criminol. (2021). https:// doi. org/ 10. 1007/ s11292- 021- 09484-9
13. Hung, T.W., Yen, C.P.: Predictive policing and algorithmic
fairness. Synthese 201(6), 206 (2023). https:// doi. org/ 10. 1007/
s11229- 023- 04189-0
14. Idowu, J., & Almasoud, A. (2023). Uncertainty in AI: Evaluat-
ing Deep Neural Networks on Out-of-Distribution Images. arXiv
preprint arXiv: 2309. 01850.
15. Ingram, E., Gursoy, F., Kakadiaris, I.A.: Accuracy, Fairness,
and Interpretability of Machine Learning Criminal Recidivism
Models. In: Ingram, E. (ed.) 2022 IEEE/ACM International Con-
ference on Big Data Computing, Applications and Technologies
(BDCAT). IEEE (2022)
16. Jain, B., Huber, M., Fegaras, L., & Elmasri, R. A. (2019, June).
Singular race models: addressing bias and accuracy in predicting
prisoner recidivism. In Proceedings of the 12th ACM Interna-
tional Conference on PErvasive Technologies Related to Assistive
Environments (pp. 599–607). ACM.
17. Khademi, A., & Honavar, V. (2020) Algorithmic bias in recidi-
vism prediction: A causal perspective (student abstract). In Pro-
ceedings of the AAAI Conference on Artificial Intelligence (Vol.
34, No. 10, pp. 13839–13840). AAAI Press. https:// doi. org/ 10.
1609/ aaai. v34i10. 7192
18. Little, R.J., Rubin, D.B.: Statistical analysis with missing data.
John. Wiley. (2019). https:// doi. org/ 10. 1002/ 97811 19482 260
19. Mohler, G., Raje, R., Carter, J., Valasik, M., Brantingham, J.: A
penalized likelihood method for balancing accuracy and fairness
in predictive policing. In: Mohler, G. (ed.) 2018 IEEE international
conference on systems, man, and cybernetics (SMC). IEEE (2018)
20. Montana, E., Nagin, D.S., Neil, R., Sampson, R.J.: Cohort bias
in predictive risk assessments of future criminal justice system
involvement. Proc. Natl. Acad. Sci. 120(23), e2301990120 (2023).
https:// doi. org/ 10. 1073/ pnas. 23019 90120
21. Page, M.J., McKenzie, J.E., Bossuyt, P.M., Boutron, I., Hoffmann,
T.C., Mulrow, C.D., Moher, D.: The PRISMA 2020 statement: an
updated guideline for reporting systematic reviews. Internat. J.
Surg. (2021). https:// doi. org/ 10. 1186/ s13643- 021- 01626-4
22. Pastaltzidis, I., Dimitriou, N., Quezada-Tavarez, K., Aidinlis, S.,
Marquenie, T., Gurzawska, A., & Tzovaras, D. (2022). Data aug-
mentation for fairness-aware machine learning: Preventing algo-
rithmic bias in law enforcement systems. In Proceedings of the
2022 ACM Conference on Fairness, Accountability, and Transpar-
ency. ACM. https:// doi. org/ 10. 1145/ 35311 46. 35346 44
23. Richardson, R., Schultz, J.M., Crawford, K.: Dirty data, bad pre-
dictions: How civil rights violations impact police data, predictive
policing systems, and justice. NYUL. Rev. Online. 94, 15 (2019)
24. Rodolfa, K. T., Salomon, E., Haynes, L., Mendieta, I. H., Lar-
son, J., Ghani, R. (2020) Case study: predictive fairness to reduce
misdemeanor recidivism through social service interventions. In
Proceedings of the 2020 Conference on Fairness, Accountability,
and Transparency. ACM. https:// doi. org/ 10. 1145/ 33510 95. 33728
63
25. Somalwar, A., Bansal, C., Lintu, N., Shah, R., Mui, P.: AI For Bias
Detection: Investigating the Existence of Racial Bias in Police
Killings. In: Somalwar, A. (ed.) 2021 IEEE MIT Undergraduate
Research Technology Conference (URTC). IEEE (2021)
26. Tripathi, R.K., Jalal, A.S., Agrawal, S.C.: Suspicious human activ-
ity recognition:a review. Artif. Intell. Rev. Intell. Rev. 50, 283–339
(2018). https:// doi. org/ 10. 1007/ s10462- 017- 9545-7
27. Udoh, E. S. (2020, September). Is the data fair? An assessment of
the data quality of algorithmic policing systems. In Proceedings
of the 13th International Conference on Theory and Practice of
Electronic Governance (pp. 1–7). ACM. https:// doi. org/ 10. 1145/
34285 02. 34285 03
28. Idowu, J. A. (2024). Debiasing Education Algorithms. Inter-
national Journal of Artificial Intelligence in Education, 1–31.
Springer Nature. https:// doi. org/ 10. 1007/ s40593- 023- 00389-4
29. Ullah, W., Ullah, A., Haq, I.U., Muhammad, K., Sajjad, M.,
Baik, S.W.: CNN features with bi-directional LSTM for real-
time anomaly detection in surveillance networks. Multimed.
Tool. App. 80, 16979–16995 (2021). https:// doi. org/ 10. 1007/
s11042- 020- 09406-3
30. Urcuqui, C., Moreno, J., Montenegro, C., Riascos, A., Dulce,
M.: Accuracy and fairness in a conditional generative adversarial
model of crime prediction. In: Urcuqui, C. (ed.) 2020 7th Interna-
tional Conference on Behavioural and Social Computing (BESC).
IEEE (2020)
31 Van Berkel, N., Goncalves, J., Hettiachchi, D., Wijenayake, S.,
Kelly, R.M., Kostakos, V.: Crowdsourcing perceptions of fair pre-
dictors for machine learning: A recidivism case study. Proceed.
ACM Hum. Comput. Int. (2019). https:// doi. org/ 10. 1145/ 33591 30
32 Yamni, M., Daoui, A., Karmouni, H., Sayyouri, M., Qjidaa, H.,
Motahhir, S., Aly, M.H.: An efficient watermarking algorithm
for digital audio data in security applications. Sci. Rep. (2023).
https:// doi. org/ 10. 1038/ s41598- 023- 45619-w
33. Yen, C.P., Hung, T.W.: Achieving equity with predictive policing
algorithms: a social safety net perspective. Sci. Eng. Eth. 27, 1–16
(2021). https:// doi. org/ 10. 1007/ s11948- 021- 00312-x
34. Gusenbauer, M., Haddaway, N.R.: Which academic search sys-
tems are suitable for systematic reviews or meta-analyses? Evalu-
ating retrieval qualities of Google Scholar, PubMed, and 26 other
resources. Res. Synth Method. 11(2), 181 (2020). https:// doi. org/
10. 1002/ jrsm. 1378
35. Angerschmid, A., Zhou, J., Theuermann, K., Chen, F., Holzinger,
A.: Fairness and explanation in AI-informed decision making.
Mach. Learn Knowl. Extr. 4(2), 556–579 (2022)
36. Posadas, B. (2017, June 26). How strategic is Chicago's "Strate-
gic subjects list"? upturn investigates. Medium. Retrieved Octo-
ber 6, 2022, from https:// medium. com/ equal- future/ how- strat
egic- is-chicagos-strategic-subjects-list-upturn-investigates-
9e5b4b235a7c
37. Van Dijk, J., Nieuwbeerta, P., & Joudo Larsen, J. (2021). Global
crime patterns: An analysis of survey data from 166 countries
around the world, 2006–2019. Journal of Quantitative Criminol-
ogy, 1–36.
38. Tucek, A. (2018). Constraining Big Brother: The Legal Deficien-
cies Surrounding Chicago's Use of the Strategic Subject List.U.
Chi. Legal F., 427.
39. Heinrichs, B.: Discrimination in the age of artificial intelligence.
AI Soc. 37(1), 143–154 (2022)
40. Follesdal, A.: John Rawls’ theory of justice as fairness. In: Phi-
losophy of Justice, pp. 311–328. Springer, Netherlands, Dordrecht
(2014)
Publisher's Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Content courtesy of Springer Nature, terms of use apply. Rights reserved.

1.
2.
3.
4.
5.
6.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:
use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
use bots or other automated methods to access the content or redirect messages
override any security feature or exclusionary protocol; or
share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at
onlineservice@springernature.com