Received 10 April 2023, accepted 11 June 2023, date of publication 14 June 2023, date of current version 20 June 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3286344

Crime Prediction Using Machine Learning and

Deep Learning: A Systematic Review
and Future Directions
VARUN MANDALAPU 1 , LAVANYA ELLURI 2 , (Member, IEEE),
PIYUSH VYAS2 , AND NIRMALYA ROY1 , (Member, IEEE)
1 Department of Information Systems, University of Maryland Baltimore County, Baltimore, MD 21250, USA
2 Subhani Department of Computer Information Systems, Texas A&M University—Central Texas, Killeen, TX 76549, USA
Corresponding authors: Varun Mandalapu ([email protected]) and Lavanya Elluri ([email protected])

ABSTRACT Predicting crime using machine learning and deep learning techniques has gained considerable
attention from researchers in recent years, focusing on identifying patterns and trends in crime occurrences.
This review paper examines over 150 articles to explore the various machine learning and deep learning
algorithms applied to predict crime. The study provides access to the datasets used for crime prediction by
researchers and analyzes prominent approaches applied in machine learning and deep learning algorithms to
predict crime, offering insights into different trends and factors related to criminal activities. Additionally,
the paper highlights potential gaps and future directions that can enhance the accuracy of crime prediction.
Finally, the comprehensive overview of research discussed in this paper on crime prediction using machine
learning and deep learning approaches serves as a valuable reference for researchers in this field. By gaining
a deeper understanding of crime prediction techniques, law enforcement agencies can develop strategies to
prevent and respond to criminal activities more effectively.

INDEX TERMS Crime prediction, crime detection, crime datasets, deep learning, machine learning, smart
policing, survey.

I. INTRODUCTION advancements in this field and contribute to future research

Crime prediction is a complex problem requiring advanced efforts.
analytical tools to effectively address the gaps in existing The field of machine learning is a subset of artificial
detection mechanisms. With the increasing availability of intelligence that uses statistical models and algorithms to
crime data and through the advancement of existing tech- analyze and make predictions based on data. On the other
nology, researchers were provided with a unique opportu- hand, deep learning methods are a subset of machine learning
nity to study and research crime detection using machine that uses artificial neural networks with multiple layers to
learning and deep learning methodologies. Based on the model complex relationships between inputs and outputs [4].
recent advances in this field [1], [2] [3], this article will Both machine learning and deep learning methodologies have
explore current trends in machine learning and deep learning the potential to be applied to the problem of crime prediction
for crime prediction and discuss how these cutting-edge in many ways [5].
technologies are being used to detect criminal activities, Machine learning algorithms have been utilized in crime
predict crime patterns, and prevent crime. Our primary prediction to analyze crime data and predict future crime
goal is to provide a comprehensive overview of recent patterns [6]. For example, algorithms like decision trees,
random forests, and support vector machines have been
trained on crime data from specific cities to predict crime
The associate editor coordinating the review of this manuscript and patterns accurately [7]. Apart from predicting crime pat-
approving it for publication was Rongbo Zhu . terns, these algorithms can provide valuable insights into

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 11, 2023 60153
 V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

crime trends and patterns. These capabilities allow for problem of crime prediction, it is vital to develop inter-
deploying resources and tactics to combat crime effectively. pretable models that can provide clear explanations of their
Additionally, machine learning algorithms can also be used predictions.
to identify correlations between crime incidents and various Moreover, the recent advancements in machine learning
environmental and demographic factors such as location, and deep learning for crime prediction show great promise in
weather, and time of day [8]. This information can be used to addressing this complex problem [12]. However, significant
develop crime prediction and prevention strategies suitable to challenges remain, and much work is still needed to realize
a given community’s specific needs. these technologies’ potential fully. This research article
Predictive policing is also a significant application of provides a comprehensive overview of recent trends in this
machine learning for crime prediction [9]. Predictive policing field and offers insights into the potential applications of
refers to using data and analytics to inform law enforcement machine learning and deep learning for crime prediction.
efforts and reduce crime. Machine learning algorithms can be By highlighting the potential of these technologies and
used to analyze crime data from a specific geographic area, the challenges that must be addressed, this research article
such as a city or neighborhood, to identify crime hotspots contributes to the broader research community. It advances
and predict future crime incidents. This information can then our understanding of the role of machine learning and deep
be used to direct policing resources to areas where they are learning in crime prediction. Hence, the key contributions
most needed, increasing the effectiveness of law enforcement of this work are as follows:- first, this paper provides the
efforts. amalgamation of existing studies that utilized state-of-the-
Deep learning algorithms, such as convolution and recur- art machine learning and deep learning-based approaches
rent neural networks, have also shown promise in crime in the realm of detecting neighborhood crime. Thereby
prediction. These algorithms have been trained on crime data extending the fathomable literature knowledge base. Second,
with either a spatial or temporal component to accurately this paper eliminates the limitation of the scarcity of potential
predict crime patterns in specific cities. For example, deep datasets availability. We have highlighted distinct publicly
learning algorithms have been used to analyze crime data, available datasets related to neighborhood crime prediction
including the time, location, and type of crime incidents [10]. that existing studies have utilized. Thereby archiving the data
This information is used to create a predictive model that can resources for future scholars. Third, this work drafted future
be used to identify potential crime hotspots and predict future research directions to eliminate the existing research gaps in
crime incidents. neighborhood crimes. Thereby reasonably providing future
Another application of deep learning in crime prediction research objectives/questions to the research community to
is computer vision and video analysis. This technology has pursue further.
been used to analyze video footage from surveillance cameras
to detect and classify criminal activities, such as vandalism, II. RESEARCH METHODOLOGY
theft, and assault [1]. The advanced deep learning models The primary research aims to find various efficient algorithms
are also integrated with drones and other aerial technologies for predicting neighborhood crimes. In our previous work [8],
to provide new opportunities to monitor and respond to we used statistical analysis to predict the crimes in Newyork
criminal activities. These algorithms have also been used city. Our paper got good attention from the researchers, so we
to analyze crime data from multiple sources, including wanted to look for the efficient machine learning and deep
crime reports, social media, and police records, providing learning approaches used in this area. We have followed
a more comprehensive view of criminal activities [11]. a systematic approach to select the papers for this review.
By automating this process, deep learning algorithms have As part of this research, we have considered the papers from
the potential to enhance the ability to identify and respond to multiple databases related to predicting crime.
crime in real-time, providing a crucial tool in the fight against For this review, we have considered all the primarily used
criminal activity. terms in the papers focused on predicting crimes. To include
Despite the promise of machine learning and deep all the possible alternative words of each term, we have used
learning for crime prediction, several challenges must be ‘‘*’’ as a wild character for IEEE and ACM databases so
addressed. One of the biggest challenges is the availability that it contains zero or more characters after the string. The
of high-quality crime data. Crime data can be difficult to main target of this review is to check for all the existing
obtain, and the available data may need to be completed research works to predict crime. In addition, we want to help
or reliable. Additionally, collecting and using crime data the research community by identifying the different datasets
is associated with privacy and ethical concerns. These used to apply the algorithms. Irrelevant studies are removed
challenges must be addressed to fully realize the potential by applying multiple filters to our search queries. We also
of machine learning and deep learning for crime prediction. selected 30 papers to be part of the main text based on
Another challenge is the interpretability of machine learning relevance and novelty, and 20 more papers are added in the
and deep learning models. These models can be challenging appendix Table 7. In this survey, we have used a combination
to understand and interpret, limiting their usefulness in of an automated and manual search shown in Figure 2. In the
decision-making. To effectively apply these models to the initial stages, we focused on using the automatic digital

60154 VOLUME 11, 2023

V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

FIGURE 1. Research paper selection methodology.

search. In the final step, we manually read the entire paper

to select a set of documents in machine learning and deep
learning areas.
Firstly, we have identified the key terms to create the
queries. We then used those keywords to construct the various
research database-related queries based on respective syntax.
Below are the queries used to explore IEEE, Science Direct,
and ACM databases.
IEEE query:
((‘‘Document Title’’: ‘‘crime*") AND (‘‘Document Title":
‘‘predic*" OR ‘‘Document Title": ‘‘detec*" OR ‘‘Document
Title": ‘‘recogni*" OR ‘‘Document Title": ‘‘machine learn-
ing" OR ‘‘Document Title":‘‘deep learning" OR ‘‘Document
Title":‘‘clustering" OR ‘‘Document Title":‘‘natural language
processing"))
Science Direct Query:
(‘‘crime") AND ( ‘‘prediction" OR ‘‘detection" OR
‘‘recognition" OR ‘‘machine learning" OR ‘‘deep learning"
OR ‘‘clustering" OR ‘‘natural language processing")
ACM Query:
‘‘query": { Fulltext:((‘‘crime*") AND ( ‘‘predic*" OR
‘‘detec*" OR ‘‘recogni*" ‘‘machine learning" OR ‘‘deep FIGURE 2. Steps involved for typical crime detection.

learning" OR ‘‘clustering" OR ‘‘natural language process-

ing")) } b) ACM (DL) Digital Library (https://dl.acm.org)
c) IEEE Xplore Access Digital Library (https://ieeexplore.
A. DATA COLLECTION ieee.org)
We focused on looking into individual research libraries After applying the automated filters as shown in Figure 1,
rather than searching in google scholar. Because google all the authors manually divided the work and read the papers
scholar will have data from all these databases, there could to select the final set. For the selection, we mainly looked
be duplicates. Below are the database library homepage links into the essential elements like the focus or objective of the
where the research works were extracted using the keywords article, datasets that authors have used, algorithms applied,
mentioned in search queries. Initially, we searched using and the accuracy rates. The focus of this survey is not only
all the metadata attributes available on each database. Next, to help the community know the various algorithms applied
we applied the filters only on the full-text papers. As we but also to let them know about the datasets they can use
noticed that the number of documents is still high, we applied to apply the novel algorithms and get the results for their
the filter on the index terms used in the article as the results research.
will be more relevant. We have more than 450 papers from all
the databases at this stage. Finally, in the last step, we applied III. LITERATURE ANALYSIS
the filter on the document title, where the total number of A. PRE-ASSESSMENT LITERATURE ANALYSIS
papers was 157. An analysis of collected literature data from the distinct
a) Science-Direct Elsevier (https://www.sciencedirect.com) research databases is essential [13] to receive information

VOLUME 11, 2023 60155

 V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

FIGURE 3. Research publication trends from 2018-2022.

FIGURE 5. Distribution of article’s citations.

FIGURE 4. Distribution of article’s page counts.

FIGURE 6. Distribution of research articles at various venues.

regarding the growth of an adopted research domain, scope

across the research community, and popularity among the To assess the popularity of the neighborhood crime
existing researchers. Thus, we have performed a detailed research articles, we have also performed the citation
analysis of the collected literature data. We did pre-analysis analysis shown in Figure 5. Wherein the conference papers
and post-analysis, wherein pre-analysis comprises the explo- have gained more citations (60) than the other journals
ration of initially collected literature ( i.e., the research (10 citations), chapter (1 citation), seminar (0 citations),
papers that were collected immediately after performing our symposium (3 citations), and workshop (o citation) articles in
search query), and post-analysis comprises the investigation the neighborhood crime area, this may be because this area is
of those study’s data that were finally selected after applying less popular among the researchers.
selection criteria. Although we have exhumed the two Figure 6, shows the distribution of neighborhood
renowned research databases, IEEE and Science Direct, crime-related articles published in various venues. Wherein
lately, it has been observed that Science direct discontinued this figure depicts the venue name (e.g., conference, journal,
the search result extraction. Hence following analytical etcetera), number of papers (i..e, in numeric value-1,7..), and
charts are based on IEEE databases findings. Figure 3 percentage of published articles (i.e., 1%, 82%,..). Here, the
shows the research publication trends in neighborhood crime majority of papers have been published in the conferences
from 2018 to 2022 (i.e., the last five years). It showed upward than the other venues like journals, workshops, symposiums,
trends from 2018 to 2021 and downward trends during 2022. and seminars. In the neighborhood crime area, 82% of the
The years 2020 and 2021 are the apexes of COVID-19, which articles were published in conferences, followed by 9% in
could be why existing researchers have utilized that time to the journal, 4% in the symposium, 3% in book chapters,1%
explore more neighborhood crime research. in seminars, and 1% in workshops.
Figure 4 indicates the distribution of article page counts
for the neighborhood crime research. The majority of articles B. POST-ASSESSMENT LITERATURE ANALYSIS
(i.e., 46) comprise five pages. This may be because many To enhance the understanding of the neighborhood crime
researchers have published their work in conferences and domain in combination with our above-cited selection cri-
symposiums rather than journals. Figure 4 depicts that teria, we have also performed the post-assessment literature
very few articles comprise more than ten pages. Since analysis because this is the crux of our literature survey to
conferences, seminars, workshops, and symposiums allowed fulfill our identified objectives. We have created a word cloud
the presentation of abstracts and short papers, this Figure 4 for Neighborhood crime-related papers to fathom further the
also depicts the 1,2,3 page-long articles. selected papers’ underlying key concepts or themes. A word

60156 VOLUME 11, 2023

V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

FIGURE 7. Word cloud on selected articles.

FIGURE 9. Distribution of neighborhood crime-related selected article’s
technique’s classes.

combination of ML and DL, 2% have used the combination

of DL and NLP, and 2% have used the combination
of ML and NLP. It has been observed that machine
learning techniques are popular in the neighborhood crime
area.
Figure 9 shows the distribution of neighborhood
crime-related selected article’s technique classes among
Science Direct and IEEE databases. As shown in Figure 9,
among the neighborhood crime-related articles, the classi-
FIGURE 8. Distribution of neighborhood crime-related selected article’s fication task is the prime focus utilizing various ML and
technique types. DL-related techniques. In the neighborhood crime domain
– 63% of articles have focused on the classification task,
29% have focused on the regression tasks, 6% have focused
cloud often called a tag cloud, is a graphic depiction of the on the clustering task, and 2% of the studies have utilized
terms that appear the most frequently in a given text. Each the combination of classification and clustering. It has been
word’s magnitude in the word cloud reflects how frequently observed that the Classification task is the prime focus of the
it appears in the text. Word clouds are frequently employed in studies on neighborhood crime.
literature reviews to swiftly pinpoint the key themes or topics Figure 10 shows the Distribution of neighborhood Crime
within a sizable body of material. Additionally, they can be related selected article’s Technique classes Versus technique
used to compare various texts and find trends and patterns in Type among Science Direct and IEEE databases. Figure 10
the data [14] answers the question- what AI Technique is used for which
Figure 7 shows the word cloud for the neighborhood technique classes (classification, clustering, and regression)?
crime-related selected articles wherein many crime-related Herein all five techniques (ML, DL, DL+NLP, ML+NLP,
key terms have emerged as trends in the existing stud- and ML+DL) have been used for the classification task
ies. ‘‘crime’’, ‘‘criminal’’, ‘‘policing’’, ‘‘enforcement’’, and whereas, for clustering and the combination of clustering
‘‘security’’ are the words that indicate the emphasis of and classification, the ML is solely used. In neighborhood
researchers on the sub-areas of crime detection. Moreover, crime articles NLP is also used for classification and
‘‘prediction’’, ‘‘algorithms’’, and ‘‘techniques’’ are the words regression tasks in addition to ML and DL. For the
that indicate the aim of the existing studies. This word cloud is classification and regression tasks ML and DL both have been
aligned with our selection criteria and objectives that further used.
validate our final set of selected articles for this literature
review study. IV. CRIME PREDICTION PROCESS & DATASETS
Figure 8 shows the Distribution of neighborhood Crime prediction using machine and deep learning involves
crime-related selected article’s technique types among several major steps as shown in Figure 11. The first step
Science Direct and IEEE databases, respectively. As shown is data collection, which involves gathering relevant data
in the Figure 8, among the neighborhood crime-related such as crime statistics, demographics, and weather patterns.
articles, machine learning (ML), the combination of machine The next step is data preprocessing, which includes cleaning
learning and deep learning (DL), and the combination of and transforming the data into a usable format. After data
machine learning and natural language processing (NLP), the preprocessing, the data is split into training and testing
combination of DL and NLP, and DL are the majorly used AI sets for model development and evaluation. The next step
technique type. In the neighborhood crime domain – 67% of is feature engineering, which involves selecting relevant
articles have used ML, 21% have used DL, 8% have used the features from the data that can be used to train the model.

VOLUME 11, 2023 60157

 V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

FIGURE 10. Distribution of neighborhood Crime related selected article’s Technique classes Versus technique Type.

FIGURE 11. Architecture flow of crime prediction.

Once the features are selected, various machine and deep relationship to the socio-economic factors of people based on
learning algorithms can be applied to the data for training and their geo-locations in the area.
prediction purposes. Finally, the trained models are evaluated Other datasets commonly used in crime detection and pre-
using various performance metrics to assess their accuracy diction research include the Los Angeles Crime Dataset, the
and effectiveness in predicting crime. The results can be used New York City (NYC) Crime Dataset, and the Philadelphia
to support decision-making in law enforcement and crime Crime Dataset. These datasets contain information on crimes
prevention efforts. reported in their respective cities and have been used to create
As shown in Table 1, there have been many datasets used in models that predict the likelihood of specific types of crimes
crime detection and prediction research articles. One example occurring in different areas. In addition to these, there are also
is the Chicago Crime Dataset, which contains data on crimes global datasets that focus on CCTV video footage, types of
reported in the Chicago area. This dataset has been used to aggression, and weapons for real-time crime predictions.
create models that predict the likelihood of specific types Overall, these datasets provide valuable information for
of crimes occurring in different areas of the city. Another researchers to build crime prediction models that could help
dataset used in crime prediction research is the London Crime law enforcement agencies prevent and respond to criminal
Dataset, which contains data on crimes reported in London activities more effectively. The location and access to datasets
city. This dataset has been used to create models that predict used by research articles surveyed in this paper are listed in
the likelihood of crimes occurring in specific areas and their Table 1.

60158 VOLUME 11, 2023

V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

TABLE 1. Crime detection papers with data science algorithms.

V. CRIME PREDICTION USING MACHINE LEARNING models can be used to analyze a wide range of features
TECHNIQUES and make predictions about crime patterns. In addition to
Traditional machine learning models have proven to be these techniques, traditional machine learning models can
effective for crime prediction. Various types of models also be used for anomaly detection and outlier analysis
such as decision trees, support vector machines, logistic in crime data. By identifying unusual patterns or outliers
regression, and random forests have been utilized to analyze in the data, law enforcement agencies can detect potential
crime data and identify patterns that can be used to predict criminal activity and take action to prevent it. In the below
criminal activity. Unlike deep learning, which relies on large sections V-1 and V-2, we discuss the latest research on using
amounts of data and complex neural networks, traditional machine learning model-based regression and classification
machine learning models require fewer data points and are for crime prediction.
easier to interpret. For example, a logistic regression model
can be used to predict the likelihood of a certain type of 1) MACHINE LEARNING BASED REGRESSION METHODS
crime occurring based on factors such as time of day, location, FOR CRIME PREDICTION
and demographics of the area. A decision tree model can be Several crime detection scenarios are predicted using regres-
used to identify the most important factors that contribute to sion techniques as shown in Table 2. Researchers mainly
the occurrence of a particular crime. Random Forest (RF) focused on prevalent crimes like motorcycle robbery, losing

VOLUME 11, 2023 60159

 V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

TABLE 2. Crime prediction using machine learning regression techniques.

property, and crimes in urban areas. Numerous factors four models, the authors found that the gradient boosting
may drive the boom in motorcycle robberies. For example, technique outperformed, proven to be the best method to
population growth and density, commuting conduct, bike predict the crime rate in the urban area.
usage, etc. These situations are problems for the police to In another study [51], authors have looked at the crimes
govern and screen regularly since it requires forecasting and in Brazil, which have increased rapidly in recent times.
probabilities of robbery in a precise term. A novel method is Numerous predictive solutions use intelligent systems to
proposed in the research study [49]; the authors created an identify when will a criminal offense will arise, which lets
application to predict motorcycle robbery with a technique to police send to those areas that are in danger. As part of
consider outside consequences using ARIMAX – TFM with their research, the authors looked into four machine-learning
a single input. The accuracy of ARIMAX is measured using approaches for identifying where a criminal offense will
Mean Absolute Percentage Error (MAPE) and Root Mean arise in Fortaleza, Brazil. Their results indicate that easy
Squared Error (RMSE), and the scores are 32.30 and 6.68. algorithms are efficient in predicting crime. Also, they have
Rapid urbanization is a compelling challenge connected seen that the Decision Tree and Bagging Regressor strategies
to city management and services. Cites with higher crime obtained quality prediction outcomes.
rates are difficult to manage public safety. To reduce crimes, As mentioned above, numerous linear models are there
new technologies are relieving police departments to access to predict crime through the correlations between urban
vast amounts of crime data to identify underlying trends metrics and crime. However, due to multicollinearity and
and patterns. These technologies have doubtlessly grown the nonGaussian distributions in urban attributes, we usually
efficient deployment of police assets within a given region tend to make controversial conclusions on these attributes to
and ultimately guide greater powerful crime prevention. predict crime. Ensemble-based machine learning algorithms
Researchers have worked on predictive models to use these can deal with such problems adequately. In the research
datasets and predict crimes. Study [23], provides a technique work [44], authors applied random forest regressor to predict
primarily based on spatial analysis and auto-regressive the crime and quantify the impact of urban attributes on
fashions to automatically locate excessive-hazard crime areas homicides. Their approach has 97% accuracy in crime
in city areas and reliably forecast crime tendencies in each prediction, and the significance of city indicators is clustered
area. Experiments are performed on real-world datasets and ranked equally. Their research identifies the rank of
gathered in New York City and Chicago. urban indicators based on their significance in predicting
Another study [50], compared multiple techniques to crime. As per their results, unemployment and illiteracy are
predict the crimes in various areas of a metropolis. This the essential variables for depicting homicides in Brazilian
research explored three predictive models: linear regression, towns.
logistic regression, and gradient boosting. The authors
utilized feature selection techniques to select essential 2) MACHINE LEARNING BASED CLASSIFICATION METHODS
predictors. By applying feature selection methods, there FOR CRIME PREDICTION
is an improvement in accuracy scored, and it helped to Traditional regression techniques can successfully check the
avoid model overfitting. After comparing the results of all variables’ significance but, they must be more reliable for

60160 VOLUME 11, 2023

V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

TABLE 3. Crime prediction using machine learning classification techniques.

crime prediction. In many research works [32], [33], [34] multi-class classification models like OVR-XGBoost and
mentioned in section V, authors have proven that machine OVO-XGBoost. As the theft datasets have different classes,
learning models effectively predict crimes. Still, they could be they have utilized the SMOTENN algorithm to process
more efficient in identifying which variables are significant and make data the dataset balanced. Their results show
in predicting crimes. We further examined the classification that OVR-XGBoost and OVO-XGBoost models’ prediction
techniques to predict different criminal incidents like analyz- accuracy is better than the baseline XGBoost models. In the
ing the criminal reports as shown in Table 3. Studying those study [54], the authors have selected 17 variables for crime
reviews for crime prediction enables regulatory authorities prediction, and the XGBoost algorithm is adopted to train
to deal with crime prevention strategies. However, collecting the prediction model. A post hoc interpretable approach,
these reviews personally and determining their crime types is Shapley additive explanation (SHAP), is used to parent
challenging. In one study [52] authors have created a novel the contribution of person variables. SHAP, a post hoc
approach, an incremental classifier that learns the new data interpretable method, is used to determine the significance
and dynamically predicts the results. In this research [52], of individual variables. Among all 17 variables used in this
they have utilized the Bi-objective Particle Swarm Optimiza- research, the percentage of the non-neighborhood population
tion technique to develop an efficient incremental classifier and the populace aged 25–44 contribute greater than different
for dynamically classifying and predicting crime reports. variables in predicting crime. The higher the ambient
Crime reports from various countries have been collected population of elderly 25–44 in the vicinity, the more public
from online newspapers to measure the performance of crimes. The authors have also validated the SHAP values
their classifier. Also, they evaluated the results manually to demonstrate each variable’s contribution to the crime
with unprejudiced police witness narrative crime reports. prediction across the experimental findings. These outcomes
They tested their approach on four datasets to measure their of the neighborhood techniques can assist the police in
model’s statistical significance. identifying the most important factors.At the same time, the
Another research [53], focuses on predicting crime using global model identifies the essential features of the entire
the XGBoost algorithm. Based on the records of theft region.
instances in H city, they developed an optimized decompo- Another research [55] focuses on predicting crime dur-
sition and fusion method based on XGBoost and applied ing or after psychiatric care. As modern threat-evaluation

VOLUME 11, 2023 60161

 V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

equipment is time-consuming to administer and offers 1) DEEP LEARNING BASED REGRESSION METHODS FOR
constrained accuracy, this research looked to expand a CRIME PREDICTION
predictive model designed to discover psychiatric patients Deep learning algorithms in regression analysis are used as a
liable to commit the crime. The authors utilized the tool for crime prediction to identify the factors most strongly
longitudinal nice of the affected Danish person registries, associated with crime and use these relationships to make
recognizing the 45.720 adult patients who had connected predictions about future crime patterns. The research articles
with the psychiatric system in 2014, of which 474 committed in this area highlight the strengths of regression in modeling
crimes leading to a forensic psychiatric treatment direction the relationship between multiple variables, including crime
after discharge. Authors have used four gadget studying data, weather data, demographic data, social media data,
models (Random Forest, Logistic Regression, XGBoost, and location data. A common theme among these research
and LightGBM) over various sociodemographic, judicial, articles shown in Table 4 are the use of regression combined
and psychiatric variables. Their model identified 47% of with deep learning techniques, such as convolution neural
future forensic psychiatric patients, making correct pre- networks, recurrent neural networks, attention mechanisms,
dictions in 57% of samples. This research demonstrates and sequential fusion models, to improve the accuracy of
how a clinically useful preliminary risk assessment is crime prediction.
achieved using machine learning classification techniques. In research focusing on theft crime prediction [26], the
Their research helps to flag possible forensic psychiatric authors use regression to model the relationship between
patients while in contact with the general psychiatric theft crime data, demographic data, and weather data.
system, which allows early intervention initiatives to be The regression model adopts two deep learning models,
activated. a Long Short-Term Memory (LSTM) network and a
Another research work [56] presents a graph-based ensem- Spatio-Temporal Graph Convolutional Network (ST-GCN),
ble classification approach for predicting crime reports better to predict the likelihood of theft crimes in urban communities.
than traditional classifiers. Crime reports are graphically The regression model can incorporate external information,
modeled to locate the maximal independent subset of such as weather data, which can influence crime patterns. The
features, and then they use decision tree classifiers on this LSTM and ST-GCN models capture the temporal and spatial
set. Extensive experiments are performed to compare the dependencies in the data, respectively. In another article [28],
overall performance of the proposed approach on numerous the authors use regression to model the relationship between
crime data sets. The developed ensemble classification model crime data, weather data, and social media data. The regres-
demonstrated better performance. Apart from predicting sion model is part of a more comprehensive, multi-module
crime, researchers [42], [43] also focused on interpreting approach that uses attention mechanisms and sequential
crime-related predictions. This will lead to a better under- fusion models to predict the likelihood of crimes. This
standing of what impacts crime detection. framework consists of four sub-modules, where the initial
two modules adopt St-BiLSTM and ATTN-LSTM to process
A. CRIME PREDICTION USING DEEP LEARNING temporal and spatial features. Finally, two fusion models
TECHNIQUES are used to abstract the data and make crime predictions on
Deep learning has become a popular method for crime Chicago and San Francisco crime datasets.
prediction in recent years. The studies included in the In another research focused on using spatiotemporal
reference research articles use a range of deep learning data [39], the authors use convolutional neural networks to
algorithms, such as convolution neural networks (CNN), develop a regression model on publicly available crime data
deep neural networks, and sentiment analysis, to analyze in Los Angeles. The regression model is part of a more exten-
various types of data, including text, images, audio, and social sive, mixed spatiotemporal neural network that is designed
media. These algorithms are capable of detecting patterns to make real-time predictions about the likelihood of crimes.
and anomalies in the data that can indicate criminal activity. The authors claim that using regression in combination with
One of the key strengths of deep learning is its ability to the diverse spatiotemporal neural network results in improved
handle large and complex datasets, making it well-suited to accuracy and real-time performance. Another research [29]
the task of crime prediction. For example, image analysis that applies crime risk prediction across different cities uses
algorithms can detect threatening objects in crime scenes regression to model the relationship between crime data and
and predict the likelihood of a crime occurring. Text mining demographic data from other cities. The regression model
techniques can be used to analyze crime-related tweets and is part of an unsupervised domain adaptation technique
make predictions about crime patterns. In addition, deep designed to predict the likelihood of crimes in new cities.
learning algorithms can detect anomalies in crime data in The authors claim that using regression in combination with
smart cities, which could indicate the presence of criminal the unsupervised domain adaptation technique results in
activity. Researchers used these techniques to tackle both improved accuracy in crime prediction. A recent research
regression and classification problems in crime prediction as article [57] applied machine learning and deep learning
detailed in the below sections. methods to crime data from Xiaogan, a medium-sized city

60162 VOLUME 11, 2023

V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

TABLE 4. Crime prediction using deep learning regression techniques.

in China, to predict crime hourly. The models use weather, data. For example, image-based data can provide detailed
holiday, time slot ID, and Day of week information to information about crime scenes, including the presence of
extract spatial dependency (distance graph, poi similarity, weapons and other objects that may indicate criminal intent.
and crime similarity). Temporal dependencies captured using Similarly, audio-based data can provide valuable insights
GRU are used to predict the number of incidents in different into the tone and context of a conversation, helping to
locations. identify potential illegal activities. Another advantage of deep
These research articles highlight the versatility of regres- learning for classification problems in crime detection is the
sion as a tool that can be integrated with other techniques ability to identify hidden patterns in the data that traditional
to enhance the performance of crime prediction models. methods may miss. For example, deep neural networks can be
Another commonality among the papers is the use of trained to analyze crime-related tweets, uncovering patterns
regression to model the relationship between crime data that indicate a potentially criminal act. The results of deep
and other variables, such as weather and demographic data, learning models in crime detection have been awe-inspiring.
to incorporate external information that may influence crime The papers reviewed under deep learning classification are
patterns. This allows for the creation of more comprehensive listed in Table 5
and accurate models of crime patterns. In summary, the five In crime-related classification, two main types of deep
research papers demonstrate the strengths of regression as learning algorithms are used: Convolutional Neural Networks
a tool for crime prediction, including its ability to model (CNN) and Recurrent Neural Networks (RNN). CNN’s
the relationship between multiple variables, its versatility are commonly used in image-based classification tasks,
in being integrated with other techniques, and its ability to including crime scene prediction. In the research article
incorporate external information that may influence crime focusing on crime scene data [59], CNNs are trained to
patterns. detect threatening objects in crime scenes, such as weapons.
This allows the model to comprehensively analyze the crime
2) DEEP LEARNING BASED CLASSIFICATION METHODS FOR scene, including the presence of things that may indicate
CRIME PREDICTION criminal intent. On the other hand, RNN’s are commonly
Deep learning algorithms are trained on large amounts of data used to study temporal patterns in data. In a research
to classify instances into various categories. This makes them article focusing on crime prediction based on behavioral
ideal for solving classification problems in crime detection. tracking [40], the authors use a combination of deep learning
Deep learning models can accurately organize criminal algorithms, including CNN and RNN, to analyze behavioral
activity and detect criminal intent by analyzing vast amounts tracking data and motion analysis data. The study shows that
of data, including images, audio, text, and social media this approach can effectively predict criminal activities, such

VOLUME 11, 2023 60163

 V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

TABLE 5. Crime prediction using deep learning classification techniques.

as theft and robbery, by analyzing patterns in the behavior such as 911 calls, is available and can provide a complete
and movements of individuals in a given area. In the research picture of the crime event.
articles [36] focusing on social media data, Artificial Neural These research articles including multiple other stud-
Networks are trained on crime-related text data to predict the ies [38], [48] highlight that deep learning algorithms,
likelihood of a crime occurring. These models analyze the including CNN and RNN, have successfully applied to
context and tone of the text data to classify patterns that may various data types for crime prediction and classification.
indicate criminal activity. These studies demonstrate the versatility of deep learning
In addition to the articles mentioned above, several algorithms in this field and provide valuable insights into the
other research studies [8], [41], [58] have also used deep factors contributing to criminal activity. By leveraging the
learning techniques for crime prediction and classification. strengths of these models, law enforcement agencies can gain
These studies demonstrate the versatility of deep learning a more comprehensive understanding of criminal activity and
algorithms in crime-related classification tasks, as they can take proactive measures to prevent crime from occurring.
be applied to a wide range of data types, including images,
text, audio, and social media data. For example, research VI. DISCUSSION AND FUTURE WORK
focusing on crime anomaly detection [60] uses deep learning The adoptions of machine and deep learning algorithms
algorithms, including Autoencoders and CNN, to analyze to predict or detect crime has shown great promise in
crime patterns in smart cities. The study shows that this addressing this complex problem. By utilizing vast datasets
approach can effectively detect unusual crime patterns, which and advanced algorithms, these technologies can potentially
may indicate the presence of criminal activity. In another improve the accuracy and effectiveness of crime prediction
study focusing on audio and text data [43], the authors use a models. However, despite the advances in this field, there are
multimodal deep learning model based on CNN and BERT still significant gaps in the current understanding of how these
that considers both audio and text data to classify crime- technologies can be effectively applied to the problem of
related events. This model is advantageous when audio data, crime prediction. In this section, we will discuss the potential

60164 VOLUME 11, 2023

V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

TABLE 6. Future Research Directions.

benefits of machine learning and deep learning algorithms for many earlier researchers took advantage of data related to
crime prediction and the future research. demographics, whether outside of crime-relevant datasets,
One of the primary advantages of machine learning and there is a need to develop algorithms that can accurately
deep learning algorithms for crime prediction is the ability handle data from multiple sources and integrate it into a single
to analyze large datasets and identify patterns in criminal predictive model.
activity or behavior. The ability of these algorithms to process Another significant area of focus should be to research
vast amounts of data, including social media and other online more on the ethical implications [72], [73] of using machine
sources [62], [63], can provide valuable insights into criminal learning and deep learning for crime prediction. As these
activities that are yet to be committed. Furthermore, deep technologies are used to predict individuals and communities,
learning algorithms like CNN and RNNs have been used it is important to ensure that they do not perpetuate existing
to analyze video footage from security cameras [64]. This biases or lead to discrimination [74], [75]. Furthermore, there
capability provides a more accurate and efficient means is a need for more research on the privacy implications
of detecting criminal activities. Another major benefit of of using these technologies for crime prediction [76], [77],
machine learning and deep learning for crime prediction [78], this included but not limited to the potential risks
is the ability to develop real-time prediction models [65]. of data breaches and the misuse of personal information.
These models can be used to analyze crime data in real-time Another significant gap in the existing research is the need
and to predict future crime incidents. This supports law for more research studies on the effectiveness of machine
enforcement agencies to act quickly if a criminal activity learning and deep learning for crime prediction in the real
is being committed. Additionally, integrating decentralized world [79]. While these technologies have shown great
machine learning algorithms with wearable technology, such promise in this area, there is a need for more rigorous
as body cameras and smartwatches [1], [66], provides new evaluations of their accuracy and effectiveness [80] in real-
opportunities to collect and analyze data related to criminal world scenarios. Additionally, there is a need for more
activities. research on the scalability of these technologies and the
Even though machine learning and deep learning algo- challenges associated with their implementation in large-
rithms support effective crime prediction, there are still scale systems.
some significant challenges that needs to be addressed. Overall, machine learning and deep learning method-
One of the major challenges in this area is the need for ologies have the potential to transform the field of crime
interpretable models [54], [67] that can provide clear expla- prediction by providing more accurate and effective methods
nations of their predictions. This is particularly important for predicting criminal activities. However, in order to fully
in the context of crime prediction, as incorrect predictions realize the potential of these technologies, it is important to
might lead to serious consequences for individuals and address the existing research gaps and challenges, including
communities [68]. Apart from the existing model-based the need for interpretable models, accurate and reliable data,
explanation methods, it is also important to incorporate causal ethical considerations, and more rigorous evaluations of
based explanations [69], [70] that focus on cause and effect their accuracy and effectiveness. By addressing these gaps,
relationship between crime patterns and relevant feature we can advance our understanding of the role of machine
variables. Another challenge that needs to be addressed is learning and deep learning algorithms in crime prediction and
the need for more accurate and reliable data. In order to contribute to the development of more effective and efficient
effectively apply machine learning and deep learning for policing strategies.
crime prediction, it is important to have access to high-quality As a future research goal and agenda, we have illustrated
and up-to-date crime data [71]. As this study showed that a range of prospective research directions in the area of

VOLUME 11, 2023 60165

 V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

TABLE 7. Crime detection papers with data science algorithms.

60166 VOLUME 11, 2023

V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

TABLE 7. (Continued.) Crime detection papers with data science algorithms.

neighborhood crime based on the importance and also VII. CONCLUSION

current lack of focus in the areas. From Table 6, future The complexity of crimes has increased along with tech-
researchers may want to address concerns like - ‘‘Are nological development, creating difficult problems for law
there any reinforcement learning techniques available to enforcement. Researchers’ interest in utilizing machine
detect the neighborhood crime?’’, and What are the visual learning and deep learning to predict crime has increased
features to be considered to detect the neighborhood crime? recently, with an emphasis on finding patterns and trends in
The relevant datasets are available for such identified crime occurrences. In order to analyze the various machine
future research questions; such data may be utilized to learning and deep learning algorithms used in predicting
accomplish the goal of early recognition of neighborhood crime, this paper looks at more than 150 articles. We have
crime. significantly studied the selected 51 articles to extract the
The presented literature base and futuristic research goals essence of utilized various ML and DL techniques along
offer a number of elements to direct future studies and with the publicly available datasets. The use of machine
thus theoretically support the effort to identify neighborhood learning and deep learning algorithms to anticipate or identify
crimes. Our systematic review offers a thorough grasp of criminal activity has shown significant promise in resolving
the characteristics and methods used by earlier research to the crime detection problem. These advances may help to
recognize and detect crimes. In addition, we have outlined increase the precision and efficacy of crime prediction models
8 more research questions that fall into the categories of by leveraging large datasets and sophisticated algorithms.
technique-oriented questions and feature-oriented questions Although there is a lack of literary wisdom on how these
in Table 6. Practically speaking, this systematic review technologies can be used to solve the problem of crime
serves as a guide for various researchers, practitioners, first prediction, despite the advancements in this sector. Thus
responders, and crime analysts, to take into account the our findings help to understand the implications of various
studied features and techniques to effectively understand and ML and DL techniques. Also, our mentioned datasets and
detect the crimes that, in part, foster the effort for early crime future directions will help the existing research community
detection. to pursue their research in the area of crime prediction.

VOLUME 11, 2023 60167

 V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

APPENDIX [20] B. Sivanagaleela and S. Rajesh, ‘‘Crime analysis and prediction using
ADDITIONAL CRIME DETECTION PAPERS USING fuzzy C-means algorithm,’’ in Proc. 3rd Int. Conf. Trends Electron.
Informat. (ICOEI), Apr. 2019, pp. 595–599.
MACHINE LEARNING AND DEEP LEARNING
[21] A. M. Shermila, A. B. Bellarmine, and N. Santiago, ‘‘Crime data analysis
See Table 7. and prediction of perpetrator identity using machine learning approach,’’
in Proc. 2nd Int. Conf. Trends Electron. Informat. (ICOEI), May 2018,
pp. 107–114.
ACKNOWLEDGMENT
[22] C. Catlett, E. Cesario, D. Talia, and A. Vinci, ‘‘A data-driven
(Varun Mandalapu, Lavanya Elluri, and Piyush Vyas con- approach for spatio-temporal crime predictions in smart cities,’’ in
tributed equally to this work.) Proc. IEEE Int. Conf. Smart Comput. (SMARTCOMP), Jun. 2018,
pp. 17–24.
[23] C. Catlett, E. Cesario, D. Talia, and A. Vinci, ‘‘Spatio-temporal crime
REFERENCES predictions in smart cities: A data-driven approach and experiments,’’
[1] N. Shah, N. Bhagat, and M. Shah, ‘‘Crime forecasting: A machine Pervas. Mobile Comput., vol. 53, pp. 62–74, Feb. 2019.
learning and computer vision approach to crime prediction and pre- [24] F. Yi, Z. Yu, F. Zhuang, X. Zhang, and H. Xiong, ‘‘An inte-
vention,’’ Vis. Comput. Ind., Biomed., Art, vol. 4, no. 1, pp. 1–14, grated model for crime prediction using temporal and spatial fac-
Apr. 2021. tors,’’ in Proc. IEEE Int. Conf. Data Mining (ICDM), Nov. 2018,
[2] S. A. Chun, V. A. Paturu, S. Yuan, R. Pathak, V. Atluri, and N. R. Adam, pp. 1386–1391.
‘‘Crime prediction model using deep neural networks,’’ in Proc. 20th Annu.
[25] S. K. Dash, I. Safro, and R. S. Srinivasamurthy, ‘‘Spatio-temporal
Int. Conf. Digit. Government Res., Jun. 2019, pp. 512–514.
prediction of crimes using network analytic approach,’’ in Proc. IEEE Int.
[3] S. S. Kshatri, D. Singh, B. Narain, S. Bhatia, M. T. Quasim, and Conf. Big Data (Big Data), Dec. 2018, pp. 1912–1917.
G. R. Sinha, ‘‘An empirical analysis of machine learning algorithms for
[26] X. Han, X. Hu, H. Wu, B. Shen, and J. Wu, ‘‘Risk prediction of theft crimes
crime prediction using stacked generalization: An ensemble approach,’’
in urban communities: An integrated model of LSTM and ST-GCN,’’ IEEE
IEEE Access, vol. 9, pp. 67488–67500, 2021.
Access, vol. 8, pp. 217222–217230, 2020.
[4] C. Janiesch, P. Zschech, and K. Heinrich, ‘‘Machine learning and deep
learning,’’ Electron. Mark., vol. 31, no. 3, pp. 685–695, Apr. 2021. [27] Z. Li, C. Huang, L. Xia, Y. Xu, and J. Pei, ‘‘Spatial-temporal hypergraph
[5] W. Safat, S. Asghar, and S. A. Gillani, ‘‘Empirical analysis for crime self-supervised learning for crime prediction,’’ in Proc. IEEE 38th Int.
prediction and forecasting using machine learning and deep learning Conf. Data Eng. (ICDE), May 2022, pp. 2984–2996.
techniques,’’ IEEE Access, vol. 9, pp. 70080–70094, 2021. [28] N. Tasnim, I. T. Imam, and M. M. A. Hashem, ‘‘A novel multi-module
[6] S. Kim, P. Joshi, P. S. Kalsi, and P. Taheri, ‘‘Crime analysis through approach to predict crime based on multivariate spatio-temporal data
machine learning,’’ in Proc. IEEE 9th Annu. Inf. Technol., Electron. Mobile using attention and sequential fusion model,’’ IEEE Access, vol. 10,
Commun. Conf. (IEMCON), Nov. 2018, pp. 415–420. pp. 48009–48030, 2022.
[7] D. M. Raza and D. B. Victor, ‘‘Data mining and region prediction based [29] B. Zhou, L. Chen, S. Zhao, S. Li, Z. Zheng, and G. Pan, ‘‘Unsu-
on crime using random forest,’’ in Proc. Int. Conf. Artif. Intell. Smart Syst. pervised domain adaptation for crime risk prediction across cities,’’
(ICAIS), Mar. 2021, pp. 980–987. IEEE Trans. Computat. Social Syst., early access, Sep. 29, 2022, doi:
[8] L. Elluri, V. Mandalapu, and N. Roy, ‘‘Developing machine learning based 10.1109/TCSS.2022.3207987.
predictive models for smart policing,’’ in Proc. IEEE Int. Conf. Smart [30] U. M. Butt, S. Letchmunan, F. H. Hassan, M. Ali, A. Baqir, T. W. Koh,
Comput. (SMARTCOMP), Jun. 2019, pp. 198–204. and H. H. R. Sherazi, ‘‘Spatio-temporal crime predictions by leveraging
[9] A. Meijer and M. Wessels, ‘‘Predictive policing: Review of benefits artificial intelligence for citizens security in smart cities,’’ IEEE Access,
and drawbacks,’’ Int. J. Public Admin., vol. 42, no. 12, pp. 1031–1039, vol. 9, pp. 47516–47529, 2021.
Sep. 2019. [31] S. Yao, M. Wei, L. Yan, C. Wang, X. Dong, F. Liu, and Y. Xiong,
[10] S. Hossain, A. Abtahee, I. Kashem, M. M. Hoque, and I. H. Sarker, ‘‘Prediction of crime hotspots based on spatial factors of random forest,’’
‘‘Crime prediction using spatio-temporal data,’’ in Computing Science, in Proc. 15th Int. Conf. Comput. Sci. Educ. (ICCSE), Aug. 2020,
Communication and Security. Gujarat, India: Springer, 2020, pp. 277–289. pp. 811–815.
[11] M. Saraiva, I. Matijosaitiene, S. Mishra, and A. Amante, ‘‘Crime [32] M. Sathiyanarayanan, A. K. Junejo, and O. Fadahunsi, ‘‘Visual analysis
prediction and monitoring in Porto, Portugal, using machine learning, of predictive policing to improve crime investigation,’’ in Proc. Int. Conf.
spatial and text analytics,’’ ISPRS Int. J. Geo-Inf., vol. 11, no. 7, p. 400, Contemp. Comput. Informat. (ICI), Dec. 2019, pp. 197–203.
Jul. 2022.
[33] A. Araujo, N. Cacho, L. Bezerra, C. Vieira, and J. Borges, ‘‘Towards a
[12] O. Kounadi, A. Ristea, A. Araujo, and M. Leitner, ‘‘A systematic review crime hotspot detection framework for patrol planning,’’ in Proc. IEEE
on spatial crime forecasting,’’ Crime Sci., vol. 9, pp. 1–22, Dec. 2020. 20th Int. Conf. High Perform. Comput. Commun., IEEE 16th Int. Conf.
[13] L. J. Morrisey, ‘‘Bibliometric and bibliographic analysis in an era of Smart City, IEEE 4th Int. Conf. Data Sci. Syst. (HPCC/SmartCity/DSS),
electronic scholarly communication,’’ in Scholarly Communication in Jun. 2018, pp. 1256–1263.
Science and Engineering Research in Higher Education. Evanston, IL,
[34] A. A. Almuhanna, M. M. Alrehili, S. H. Alsubhi, and L. Syed, ‘‘Prediction
USA: Routledge, 2013, pp. 149–160.
of crime in neighbourhoods of New York City using spatial data analysis,’’
[14] M. Hofmann and A. Chisholm, Text Mining and Visualization: Case
in Proc. 1st Int. Conf. Artif. Intell. Data Analytics (CAIDA), Apr. 2021,
Studies Using Open-Source Tools, vol. 40. Boca Raton, FL, USA: CRC
pp. 23–30.
Press, 2016.
[15] P. Tamilarasi and R. U. Rani, ‘‘Diagnosis of crime rate against women [35] A. Baqir, S. U. Rehman, S. Malik, F. U. Mustafa, and U. Ahmad,
using K-fold cross validation through machine learning,’’ in Proc. ‘‘Evaluating the performance of hierarchical clustering algorithms to detect
4th Int. Conf. Comput. Methodologies Commun. (ICCMC), Mar. 2020, spatio-temporal crime hot-spots,’’ in Proc. 3rd Int. Conf. Comput., Math.
pp. 1034–1038. Eng. Technol. (iCoMET), Jan. 2020, pp. 1–5.
[16] A. Kumar, A. Verma, G. Shinde, Y. Sukhdeve, and N. Lal, ‘‘Crime [36] A. Algefes, N. Aldossari, F. Masmoudi, and E. Kariri, ‘‘A text-mining
prediction using K-nearest neighboring algorithm,’’ in Proc. Int. Conf. approach for crime tweets in Saudi Arabia: From analysis to prediction,’’
Emerg. Trends Inf. Technol. Eng., Feb. 2020, pp. 1–4. in Proc. 7th Int. Conf. Data Sci. Mach. Learn. Appl. (CDMA), Mar. 2022,
[17] S. Agarwal, L. Yadav, and M. K. Thakur, ‘‘Crime prediction based on pp. 109–114.
statistical models,’’ in Proc. 11th Int. Conf. Contemp. Comput. (IC), [37] S. P. C. W. Sandagiri, B. T. G. S. Kumara, and B. Kuhaneswaran,
Aug. 2018, pp. 1–3. ‘‘Detecting crime related Twitter posts using artificial neural networks
[18] S. R. Bandekar and C. Vijayalakshmi, ‘‘Design and analysis of machine based approach,’’ in Proc. 20th Int. Conf. Adv. ICT Emerg. Regions (ICTer),
learning algorithms for the reduction of crime rates in India,’’ Proc. Nov. 2020, pp. 5–10.
Comput. Sci., vol. 172, pp. 122–127, Jan. 2020. [38] M. A. Permana, M. I. Thohir, T. Mantoro, and M. A. Ayu, ‘‘Crime rate
[19] A. Gahalot, S. Dhiman, and L. Chouhan, ‘‘Crime prediction and detection based on text mining on social media using logistic regression
analysis,’’ in Proc. 2nd Int. Conf. Data, Eng. Appl. (IDEA), Feb. 2020, algorithm,’’ in Proc. IEEE 7th Int. Conf. Comput., Eng. Design (ICCED),
pp. 1–6. Aug. 2021, pp. 1–6.

60168 VOLUME 11, 2023

V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

[39] X. Zhou, X. Wang, G. Brown, C. Wang, and P. Chin, ‘‘Mixed spatio- [60] S. Chackravarthy, S. Schmitt, and L. Yang, ‘‘Intelligent crime anomaly
temporal neural networks on real-time prediction of crimes,’’ in Proc. 20th detection in smart cities using deep learning,’’ in Proc. IEEE 4th Int. Conf.
IEEE Int. Conf. Mach. Learn. Appl. (ICMLA), Dec. 2021, pp. 1749–1754. Collaboration Internet Comput. (CIC), Oct. 2018, pp. 399–404.
[40] R. Shenoy, D. Yadav, H. Lakhotiya, and J. Sisodia, ‘‘An intelligent [61] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and
framework for crime prediction using behavioural tracking and motion L. Fei-Fei, ‘‘Large-scale video classification with convolutional neural
analysis,’’ in Proc. Int. Conf. Emerg. Smart Comput. Informat. (ESCI), networks,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., Jun. 2014,
Mar. 2022, pp. 1–6. pp. 1725–1732.
[41] N. Aldossari, A. Algefes, F. Masmoudi, and E. Kariri, ‘‘Data science [62] D. Yang, T. Heaney, A. Tonon, L. Wang, and P. Cudré-Mauroux,
approach for crime analysis and prediction: Saudi Arabia use-case,’’ in ‘‘CrimeTelescope: Crime hotspot prediction based on urban and social
Proc. 5th Int. Conf. Women Data Sci. Prince Sultan Univ. (WiDS PSU), media data fusion,’’ World Wide Web, vol. 21, no. 5, pp. 1323–1347,
Mar. 2022, pp. 20–25. Sep. 2018.
[63] A. Ristea, M. Al Boni, B. Resch, M. S. Gerber, and M. Leitner, ‘‘Spatial
[42] Y. Ma, K. Nakamura, E. Lee, and S. S. Bhattacharyya, ‘‘EADTC:
crime distribution and prediction for sporting events using social media,’’
An approach to interpretable and accurate crime prediction,’’ in Proc. IEEE
Int. J. Geographical Inf. Sci., vol. 34, no. 9, pp. 1708–1739, Sep. 2020.
Int. Conf. Syst., Man, Cybern. (SMC), Oct. 2022, pp. 170–177.
[64] M. Muthamizharasan and R. Ponnusamy, ‘‘Forecasting crime event
[43] M. Boukabous and M. Azizi, ‘‘Multimodal sentiment analysis using audio rate with a CNN-LSTM model,’’ in Innovative Data Communica-
and text for crime detection,’’ in Proc. 2nd Int. Conf. Innov. Res. Appl. Sci., tion Technologies and Application. Berlin, Germany: Springer, 2022,
Eng. Technol. (IRASET), Mar. 2022, pp. 1–5. pp. 461–470.
[44] L. G. A. Alves, H. V. Ribeiro, and F. A. Rodrigues, ‘‘Crime prediction [65] B. Wang, P. Yin, A. L. Bertozzi, P. J. Brantingham, S. J. Osher, and J. Xin,
through urban metrics and statistical learning,’’ Phys. A, Stat. Mech. Appl., ‘‘Deep learning for real-time crime forecasting and its ternarization,’’ Chin.
vol. 505, pp. 435–443, Sep. 2018. Ann. Math., Ser. B, vol. 40, no. 6, pp. 949–966, Nov. 2019.
[45] J. He and H. Zheng, ‘‘Prediction of crime rate in urban neighborhoods [66] P. William, A. Shrivastava, N. S. Karpagam, T. Mohanaprakash,
based on machine learning,’’ Eng. Appl. Artif. Intell., vol. 106, Nov. 2021, K. Tongkachok, and K. Kumar, ‘‘Crime analysis using computer vision
Art. no. 104460. approach with machine learning,’’ in Mobile Radio Communications and
[46] H. K. R. ToppiReddy, B. Saini, and G. Mahajan, ‘‘Crime prediction & 5G Networks. Berlin, Germany: Springer, 2023, pp. 297–315.
monitoring framework based on spatial analysis,’’ Proc. Comput. Sci., [67] C. Wang, B. Han, B. Patel, and C. Rudin, ‘‘In pursuit of interpretable,
vol. 132, pp. 696–705, Jan. 2018. fair and accurate machine learning for criminal recidivism prediction,’’
[47] A. Wolf, T. R. Fanshawe, A. Sariaslan, R. Cornish, H. Larsson, and J. Quant. Criminol., vol. 39, pp. 519–581, Mar. 2022.
S. Fazel, ‘‘Prediction of violent crime on discharge from secure psychiatric [68] J. Dressel and H. Farid, ‘‘The dangers of risk prediction in the criminal
hospitals: A clinical prediction rule (FoVOx),’’ Eur. Psychiatry, vol. 47, justice system,’’ MIT Case Stud. Social Ethical Responsibilities Comput.,
pp. 88–93, Jan. 2018. Winter 2021, doi: 10.21428/2c646de5.f5896f9f.
[48] K. B. Sahay, B. Balachander, B. Jagadeesh, G. A. Kumar, R. Kumar, and [69] R. Moraffah, M. Karami, R. Guo, A. Raglin, and H. Liu, ‘‘Causal
L. R. Parvathy, ‘‘A real time crime scene intelligent video surveillance interpretability for machine learning—Problems, methods and evalu-
systems in violence detection framework using deep learning techniques,’’ ation,’’ ACM SIGKDD Explor. Newslett., vol. 22, no. 1, pp. 18–33,
Comput. Electr. Eng., vol. 103, Oct. 2022, Art. no. 108319. May 2020.
[70] E. Carter, T. Ward, and A. Strauss-Hughes, ‘‘The classification of crime
[49] P. E. P. Utomo, ‘‘Prediction the crime motorcycles of theft using ARIMAX-
and its related problems: A pluralistic approach,’’ Aggression Violent
TFM with single input,’’ in Proc. 3rd Int. Conf. Informat. Comput. (ICIC),
Behav., vol. 59, Jun. 2020, Art. no. 101440.
Oct. 2018, pp. 1–7.
[71] R. Richardson, J. M. Schultz, and K. Crawford, ‘‘Dirty data, bad
[50] V. Ingilevich and S. Ivanov, ‘‘Crime rate prediction in the urban envi- predictions: How civil rights violations impact police data, predictive
ronment using social factors,’’ Proc. Comput. Sci., vol. 136, pp. 472–478, policing systems, and justice,’’ NYUL Rev. Online, vol. 94, p. 15,
Jan. 2018. Jan. 2019.
[51] A. R. C. da Silva, I. C. D. P. Junior, T. L. C. da Silva, J. A. F. de Macedo, [72] P. M. Asaro, ‘‘AI ethics in predictive policing: From models of threat to
and W. C. P. Silva, ‘‘Prediction of crime location in a Brazilian city using an ethics of care,’’ IEEE Technol. Soc. Mag., vol. 38, no. 2, pp. 40–53,
regression techniques,’’ in Proc. IEEE 32nd Int. Conf. Tools Artif. Intell. Jun. 2019.
(ICTAI), Nov. 2020, pp. 331–336. [73] O. J. Gstrein, A. Bunnik, and A. Zwitter, ‘‘Ethical, legal and social
[52] P. Das, A. K. Das, J. Nayak, D. Pelusi, and W. Ding, ‘‘Incremental classifier challenges of predictive policing,’’ Catolica Law Rev., Direito Penal, vol. 3,
in crime prediction using bi-objective particle swarm optimization,’’ Inf. no. 3, pp. 77–98, 2019.
Sci., vol. 562, pp. 279–303, Jul. 2021. [74] K. Alikhademi, E. Drobina, D. Prioleau, B. Richardson, D. Purves, and
[53] Z. Yan, H. Chen, X. Dong, K. Zhou, and Z. Xu, ‘‘Research on prediction J. E. Gilbert, ‘‘A review of predictive policing from the perspective of
of multi-class theft crimes by an optimized decomposition and fusion fairness,’’ Artif. Intell. Law, vol. 30, no. 1, pp. 1–17, Mar. 2022.
method based on XGBoost,’’ Exp. Syst. Appl., vol. 207, Nov. 2022, [75] M. Tonry, ‘‘Predictions of dangerousness in sentencing: Déjà vu all over
Art. no. 117943. again,’’ Crime Justice, vol. 48, pp. 439–482, May 2019.
[54] X. Zhang, L. Liu, M. Lan, G. Song, L. Xiao, and J. Chen, ‘‘Interpretable [76] R. Muhlhoff, ‘‘Predictive privacy: Towards an applied ethics of data
machine learning models for crime prediction,’’ Comput., Environ. Urban analytics,’’ Ethics Inf. Technol., vol. 23, no. 4, pp. 675–690, Dec. 2021.
Syst., vol. 94, Jun. 2022, Art. no. 101789. [77] T.-W. Hung and C.-P. Yen, ‘‘On the person-based predictive policing of
[55] M. L. Trinhammer, A. C. H. Merrild, J. F. Lotz, and G. Makransky, AI,’’ Ethics Inf. Technol., vol. 23, no. 3, pp. 165–176, Sep. 2021.
‘‘Predicting crime during or after psychiatric care: Evaluating machine [78] D. Leslie, ‘‘Understanding bias in facial recognition technologies,’’ 2020,
learning for risk assessment using the Danish patient registries,’’ J. arXiv:2010.07023.
Psychiatric Res., vol. 152, pp. 194–200, Aug. 2022. [79] I. H. Sarker, ‘‘Machine learning: Algorithms, real-world applications and
research directions,’’ Social Netw. Comput. Sci., vol. 2, no. 3, p. 160,
[56] A. K. Das and P. Das, ‘‘Graph based ensemble classification for
May 2021.
crime report prediction,’’ Appl. Soft Comput., vol. 125, Aug. 2022,
[80] S. Goel, R. Shroff, J. Skeem, and C. Slobogin, ‘‘The accuracy, equity, and
Art. no. 109215.
jurisprudence of criminal risk assessment,’’ in Research Handbook on Big
[57] W. Liang, Y. Wang, H. Tao, and J. Cao, ‘‘Towards hour-level crime Data Law. Cheltenham, U.K.: Edward Elgar Publishing, 2021, pp. 9–28.
prediction: A neural attentive framework with spatial–temporal-categorical [81] R. Yadav and S. K. Sheoran, ‘‘Crime prediction using auto regression
fusion,’’ Neurocomputing, vol. 486, pp. 286–297, May 2022. techniques for time series data,’’ in Proc. 3rd Int. Conf. Workshops Recent
[58] U. V. Navalgund and K. Priyadharshini, ‘‘Crime intention detection system Adv. Innov. Eng. (ICRAIE), Nov. 2018, pp. 1–5.
using deep learning,’’ in Proc. Int. Conf. Circuits Syst. Digit. Enterprise [82] H. Wang and S. Ma, ‘‘Preventing crimes against public health with artificial
Technol. (ICCSDET), Dec. 2018, pp. 1–6. intelligence and machine learning capabilities,’’ Socio-Economic Planning
[59] M. Nakib, R. T. Khan, Md. S. Hasan, and J. Uddin, ‘‘Crime scene Sci., vol. 80, Mar. 2022, Art. no. 101043.
prediction by detecting threatening objects using convolutional neural [83] J. Abraham, R. Ng, M. Morelato, M. Tahtouh, and C. Roux, ‘‘Automati-
network,’’ in Proc. Int. Conf. Comput., Commun., Chem., Mater. Electron. cally classifying crime scene images using machine learning methodolo-
Eng., Feb. 2018, pp. 1–4. gies,’’ Forensic Sci. Int., Digit. Invest., vol. 39, Dec. 2021, Art. no. 301273.

VOLUME 11, 2023 60169

 V. Mandalapu et al.: Crime Prediction Using ML and DL: A Systematic Review and Future Directions

VARUN MANDALAPU received the master’s PIYUSH VYAS received the Bachelor of Engineer-
degree in management information systems from ing and Master of Engineering degrees in informa-
the University of Illinois Springfield, the master’s tion technology from State Technical University
degree in sensor systems technology from the Bhopal, India, in 2009 and 2012, respectively,
Vellore Institute of Technology, Vellore, India, and the M.S. and Ph.D. degrees in information
and the Ph.D. degree in artificial intelligence systems from Dakota State University, Madison,
and knowledge management from the Department SD, USA, in 2020 and 2022, respectively. He is
of Information Systems, University of Maryland currently an Assistant Professor in computer infor-
Baltimore County. He is a Research Assistant mation systems with Texas A&M University—
with the Sensor Accelerated Intelligent Learning Central Texas. He has published his articles
(SAIL) Laboratory at UMBC under Dr. Jiaqi Gong and co-advised by in IEEE TRANSACTIONS ON TECHNOLOGY AND SOCIETY, Special Issues of
Dr. Zhiyuan Chen and Dr. Karen Chen. He is also an affiliate member of Information Systems, International Journal of Information Security and
the IEEE EMBS Technical Committee on Wearable Biomedical Sensors and Privacy (IJISP-IGI Global)-ABDC, AIS-AMCIS, AIS-MWAIS, DSI, and
Systems. His research publications appeared in reputed AI venues, such as IEEE conferences. His teaching interests include machine learning, data
AAAI Workshops, Artificial Intelligence in Education, Educational Data communications, computer networks, business analytics, system analysis
Mining, Smart Health, IEEE Body Sensor Networks, and IEEE Biomedical and design, and database management systems. His current research interests
Health Informatics. include text mining, association rule mining, traditional and online machine
learning, transfer learning, deep learning, the data mining in the domain
of e-commerce, social media, micro-blogging, healthcare, medical imaging,
dark web, and explainable/conversational AI. He has received the Best
Research Paper Award for AMCIS 2021.

LAVANYA ELLURI (Member, IEEE) received NIRMALYA ROY (Member, IEEE) received the
the Ph.D. degree in information systems from bachelor’s degree in computer science and engi-
the University of Maryland Baltimore County neering from Jadavpur University, India, in 2001,
and the Master of Science degree in management and the M.S. and Ph.D. degrees in computer
information systems from the University of Hous- science and engineering from The University of
ton Clear Lake. She has worked for over a decade Texas at Arlington, in 2004 and 2008, respectively.
in the IT industry at reputed companies Infosys He is currently a Professor with the Department
and REI Systems. She has led several projects at of Information Systems, University of Maryland
REI systems and has extensive work experience Baltimore County. He directs the Mobile, Perva-
with various databases and data warehousing sive and Sensor Computing (MPSC) Laboratory.
technologies. Also, she has experience in working with a wide range of data He was a Clinical Assistant Professor with the School of Electrical
science and data analytics projects. She is currently an Assistant Professor in Engineering and Computer Science, Washington State University, from
computer information systems with Texas A&M University—Central Texas. January 2012 to June 2013. Prior to that, he was a Research Scientist with
Her research and teaching interests include data analytics, data science, the Institute for Infocomm Research (I2R), Singapore, from 2010 to 2011.
semantic web, database systems, data privacy and security, text mining, and He was a Postdoctoral Fellow with the Electrical and Computer Engineering
healthcare IT systems. Her research publications appeared in reputed venues, Department, The University of Texas at Austin, from 2008 to 2009.
such as IEEE Big Data, IEEE Cloud, IEEE ACCESS, and Frontiers in Bigdata.

60170 VOLUME 11, 2023