SUPERVISED AND UNSUPERVISED MACHINE LEARNING METHODOLOGIES FOR CRIME PATTERN ANALYSIS

International Journal of Artificial Intelligence and Applications (IJAIA), Vol.12, No.1, January 2021
DOI: 10.5121/ijaia.2021.12106 83
SUPERVISED AND UNSUPERVISED
MACHINE LEARNING METHODOLOGIES
FOR CRIME PATTERN ANALYSIS
Divya Sardana1
, Shruti Marwaha2
and Raj Bhatnagar3
1
Teradata Corp., Santa Clara, CA 95054, USA
2
Stanford University, Palo Alto, CA 94305, USA
3
University of Cincinnati, Cincinnati, OH 45219, USA
ABSTRACT
Crime is a grave problem that affects all countries in the world. The level of crime in a country has a big
impact on its economic growth and quality of life of citizens. In this paper, we provide a survey of trends of
supervised and unsupervised machine learning methods used for crime pattern analysis. We use a spatio-
temporal dataset of crimes in San Francisco, CA to demonstrate some of these strategies for crime
analysis. We use classification models, namely, Logistic Regression, Random Forest, Gradient Boosting
and Naive Bayes to predict crime types such as Larceny, Theft, etc. and propose model optimization
strategies. Further, we use a graph based unsupervised machine learning technique called core periphery
structures to analyze how crime behavior evolves over time. These methods can be generalized to use for
different counties and can be greatly helpful in planning police task forces for law enforcement and crime
prevention.
KEYWORDS
Crime pattern analysis, Machine Learning, Supervised Models, Unsupervised methods
1. INTRODUCTION
Crime is a major problem that affects all parts of the world. There are many different factors that
have an effect on the rate of crime, such as education, income level, location, time of the year,
climate, and employment conditions [1]. There are different types of crime such as vehicle theft,
bribery, extortion, terrorism etc. Depending upon the nature of crime, each type has its very own
specific characteristics which need to be investigated for mitigating it. Further, some types of
crimes exhibit similarities in their nature, such as similarities in the time of the day they occur, or
similarities in the specific location where they occur. A study of these crime characteristics as
well as similarities can be immensely helpful to the law enforcement agencies to develop a better
understanding of crimes and the factors which can help in resolving these crimes and controlling
their frequency of occurrence.
Many counties have made their crime incident reports freely available online these days [2]. This
has led to a rise of interest in the research community to discover methodologies to aid in crime
pattern analysis. Several data mining and statistical analysis techniques have been researched as
well as utilized by the law enforcement agencies to help in the identification and investigation of
crimes to make cities and counties safe.
The process of crime analysis using Machine Learning involves the use of both supervised as
well as unsupervised models to gain insights from both structured and unstructured data. In this

84
paper we classify the currently available machine learning techniques to analyze crimes as
belonging to supervised or unsupervised methodologies. Next, we demonstrate the use of both
supervised as well as unsupervised machine learning methodologies to analyze a spatio-temporal
dataset of crimes in San Francisco (SF).
For supervised learning, we use classification techniques, namely, Logistic Regression, Random
Forest, Gradient Boosting classifier and Naive Bayes model to predict the type of crime based
upon features such as crime location and time. We illustrate the process we use in data cleaning,
feature extraction, model building and evaluation. Furthermore, we improve our model
performance by accumulating data into three super classes, namely, infractions, misdemeanors
and felonies.
For unsupervised learning, we use an unsupervised graph algorithm to find core periphery
structures to analyze SF crime data. Specifically, in a temporal dataset of crimes, core periphery
structures help us to study relationships between very dense nodes which lie in core clusters
surrounded by sparse periphery nodes. Further, the way these relationships change over time
reveal many interesting patterns in the crime datasets. We demonstrate use cases where core
periphery structures are a better suited unsupervised machine learning methodology over
clustering for analyzing patters in dense graphs originating from crime datasets.
The rest of the paper is organized as follows. In section 2, we provide a survey of trends in
supervised and unsupervised machine learning techniques used in literature for the analysis of
crime datasets. In section 3, we provide a description of the SF crime dataset. Next, we outline
the supervised learning approaches that we have used to analyze SF crime dataset in section 4.
Here, we describe our methodology used along with evaluation results and suggested
optimization techniques. In section 5, we describe the use of an unsupervised machine learning
methodology, namely, core periphery structures to help understand the evolution of crimes over
years. Finally, we provide a conclusion of our paper in section 6.
2. SURVEY OF TRENDS OF SUPERVISED AND UNSUPERVISED MACHINE
LEARNING ALGORITHMS FOR CRIME ANALYSIS
In literature, many machine learning strategies have been used to analyze crime data with case
studies using different crime datasets. In this section we provide a survey of trends of machine
learning approaches used for crime analysis. We also provide a brief summary of the machine
learning methods that we implement for the analysis of SF crime dataset.
Crimes can be classified into different categories, such as, violent crimes (e.g., murder), traffic
violence, sexual assault and cyber-crimes. Depending upon the nature of the crime, different
machine learning technologies are suitable to study those crimes [2]. Both supervised as well
unsupervised machine learning methods have been used in literature for the analysis of crime
datasets.
First, we provide a survey of supervised machine learning methods that have been used in
literature for crime analysis. In [3], a crime hotspot prediction algorithm was developed using
Linear Discriminant Analysis (LDA) and K Nearest Neighbor (KNN). Cesario et al. [4]
developed an Auto-Regressive Integrative Moving Average model (ARIMA) to build a predictive
model for crime trend forecasting in urban populations. In [5], the authors used two classification
algorithms, namely, K-nearest Neighbor (KNN) and boosted decision tree to analyze a
Vancouver Police department crime dataset. In [6], Edoka used different classification models,
such as Logistic Regression, K Nearest Neighbors and XGBoost to classify the types of crimes in

85
dataset of crimes from Chicago crime porter. In [7], the authors build an ensemble learning model
to predict spatial occurrences of crimes of different types. Cichosz [8] used point of interest-
based data (such as bus stops, cinema halls, etc.) from geographical information systems to build
a crime risk prediction model for urban areas.
In this paper, we use four classification methodologies, namely, logistic regression, Random
Forest, Gradient Boosting classifier and Naive Bayes model to predict the crime category in SF
crime dataset. Logistic regression and Naïve Bayes algorithms are used to build baseline models.
Next, ensemble learning based models, Gradient Boosting and Random Forest are used to build
models which combine output of multiple classifiers. Such classifiers are known to reduce
variance of the final model and require fine parameter tuning [9]. We provide a comparative
analysis of these models using evaluation measures called F1 Measure and logloss ratio [10]. We
prefer these measures over accuracy as our evaluation metric because of the imbalanced nature of
class distribution in the dataset. Further, we suggest optimizations in modeling process to
improve the achieved accuracy. The prediction of crime types using classification techniques for
SF crime dataset has been studied before in [11], [12], [13], [14], [15] and [16]. Our classification
approach is unique in the way we extract the feature zip code used for model building from
latitude, longitude. Zip code combines two features into one and is much easier to interpret for
county police officers. Further, we suggest modelling optimizations by grouping crime types into
three super classes (Infractions, Misdemeanors and Felonies). This is a standard crime grouping
used in criminal law [17].
Next, we provide a survey of unsupervised methodologies used in literature for crime pattern
analysis. Sharma et al. [18] used K-means clustering to identify patterns in cyber-crime. They
used text mining-based approaches to transform data from web pages into features to be used for
clustering. Kumar and Toshniwal [19] used K-modes clustering algorithm to group road
accidents occurring in Dehradun (India) into segments. They further used association rule mining
to link the cause of accidents along with the identified clusters. Joshi et al. [20] used K-means
clustering on a dataset of crimes from New South Wales, Australia to identify cities with high
crime rates. In [21], the authors used fuzzy c-means algorithm to identify potential crime
locations for different cognizable crimes such as burglary, robbery and theft.
While different clustering-based approaches have been used to analyze crime patterns in
literature, another unsupervised technique called core periphery structures has not yet been fully
utilized to analyze crime datasets. Core-periphery structures are suited to applications where the
entities and relationships in a crime dataset can be expressed as a network or a graph. It helps to
extract very dense core clusters surrounded by sparse periphery clusters. For example, in crime
type corruption, a network of all emails exchanged between the involved entities could be studied
using core periphery structures to find out the chief (core) players involved in corruption
surrounded by their cohorts.
In literature, it has been demonstrated that in many crime networks, core periphery structures
naturally exist, consisting of a core of nodes densely connected to one another and a surrounding
sparse periphery [22]. In [23], Xu and Chen provide a topological analysis of a terrorist network
to show the presence of a core group consisting of Osama Bin Laden and his closest personnel
who issue orders to people in the rest of the network. In [24], the authors analyzed the structure
of a drug trafficking mafia organization in Southern Calabria, Italy. Using graph centrality
measures, they identified a few most central players that had a more active role in criminal
activities than other subjects in the network. A core periphery-based model is used in [25] to
study a Czech political corruption network. The authors demonstrate different types of ties
between network entities depending upon their position in the core or periphery groups. In [26],

86
the authors provide a core periphery analysis of the email network of people involved in the
corruption activity that led to the collapse of Enron Corporation.
Several other network-based crime analysis techniques have been proposed in literature. In [27]
the authors use a graph-based link prediction algorithm to identify missing links in a criminal
network constructed using a dataset for an Italian criminal case against a mafia. Budur et al. [28]
developed a hybrid approach using Gradient Boosted Machine (GBM) models as well as a
weighted pagerank model to identify hidden links in crime-based networks.
In this paper, we illustrate the importance of core periphery structures in identifying patterns in
criminal networks that vary over time and space. We further demonstrate that in some situations
core periphery structures are better suited than clustering algorithms as the choice of
unsupervised learning technique in the investigation of crime datasets.
In a nutshell, as a trend we see that supervised learning approaches are more suited for tasks such
as prediction of crime type, prediction of time or geographical location of crimes where crimes
come from multiple crime types. The unsupervised learning approaches are more applicable to
crime types where actors or players are involved with explicit links or relationships amongst
them, such as email or verbal exchanges between them. Examples include crime types such as
corruption, terrorism, fraud, prostitution and kidnapping.
3. SAN FRANCISCO CRIME DATASET
We use a dataset of crimes that took place in San Francisco, CA from 2003 to 2015. The police
department of San Francisco has made its crime records data public over at [29] Further, a part of
the city’s crime data was made available as a Kaggle dataset [30] as part of an open competition
to predict crime types occurring at different places in the city. The original source of this dataset
is from San Francisco’s Open Data platform [31].
3.1. Dataset Description and Feature Extraction
The SF crime dataset varies spatially as well as temporally. The crimes that occurred in different
geographical locations of SF between the years 2003 to 2015 are documented in the dataset, with
details about the hour and day of the week of occurrence. Each crime is annotated with a crime
category, such as Embezzlement and Larceny. There are 878,049 incidents of crime or total rows
in the dataset. A snapshot of this dataset is provided in figure 1. Each row in the dataset
comprises of columns: timestamp, day of the week, description of crime, latitude, longitude,
crime type, police district, how the incident was resolved and address of the incident. Using the
timestamp feature Dates, we further extracted the derived features, hour, month and year for each
row. These are easier to interpret in the modelling algorithm. Further, we used the latitude and
longitude information to determine the zip code corresponding to each crime incident. For this
task, we used the ZipCodeDatabase from pyzipcode [32] python package. This process of feature
extraction is visualized in figure 2.

87
Figure 1. Snapshot of SF crime dataset
Figure 2. Feature Extraction Process from the SF Crime dataset
In figure 3, we visually present the 39 crime categories that occur in the dataset. In figures 4, 5
and 6, we demonstrate the spatial (over zip code) as well as temporal (over years and over hours)
variation in the SF crime dataset.
Given the spatial and temporal variation in the dataset, we decided to use the features Hour,
Month, Year, Days of Week, Police District, Zip code for our supervised and unsupervised
learning tasks.

88
Figure 3. 39 different crime categories in the SF crime dataset
Figure 4. Spatial variation (over zip codes) in the SF crime dataset
Figure 5. Temporal variation (over years) in the SF crime dataset

89
Figure 6. Temporal variation (over hours) in the SF crime dataset
4. SUPERVISED LEARNING METHODOLOGIES TO ANALYZE SF CRIME
DATASET
A major challenge in using supervised machine learning techniques for crime prediction is to
build scalable and efficient models that make accurate and timely predictions. Keeping this in
mind, we developed different classification models to train the SF crime dataset and performed a
comparative analysis of these models to choose a model that best fits the analysis task.
For all the models built, we used the features Hour, Month, Year, Days of Week, Police District,
and Zip code. This enables us to capture both the spatial as well as temporal variation present in
the dataset. The goal of modelling is to predict the crime type, the column “Category” in the input
dataset as the target variable used in the model. Next, we describe our modelling approaches and
methodology in detail.
4.1. Model Building
We used four classification approaches for our model building task, namely, Logistic Regression,
Random Forest, Gradient Boosting and Naive Bayes classifier [9]. In figure 7, we describe the
whole workflow that we used in the modelling process. We started our analysis with data
exploration and feature selection as described in section 3. We converted all categorical variables,
such as “days of the week” to dummy variables using one-hot encoding. Next, we performed 10-
fold cross validation on the training set, simultaneously doing parameter tuning for the input
models. All the classification algorithms treat this modelling task as a multi class classification
problem. First, we built baseline models using Naïve Bayes and Logistic Regression. Next, we
moved on to building ensemble learning models using Random Forest and Gradient Boosting
approaches. These are more powerful modelling techniques which combine multiple classifiers to
generate output and lead to variance minimization. These two algorithms require some intricate
parameter tuning. We tuned two parameters namely, “max tree depth” and “number of trees” by
trying different permutations of the parameters and choosing the ones which gave the best value
of F1-measure and logloss. These measures are described in detail in section 4.2. The developed
models were then used to get predictions for the validation set. The models built were finally
evaluated using the measures described in the next section.

90
Figure 7. Workflow used in Model Building task for SF crime category classification
4.2. Model Evaluation
For evaluating the models built, we used two evaluation metrics, namely, F1 measure and
Logloss [10]. Let TP, FP, TN, FN denote the True Positives, False Positives, True Negatives and
False Negatives respectively in the confusion matrix constructed for the classification problem.
The formulation for the evaluation measure Accuracy is as below.
For an imbalanced dataset, like the current SF crime dataset with an imbalanced class
distribution, if we output all data points to belong to the majority class, then accuracy results will
show a misleading value of high accuracy. For this reason, we do not use accuracy evaluation
measure for our case. We use F1 Measure as formulated below.
F1 measure uses a harmonic mean of precision and recall and ensures that both are balanced in
the output. A high F1 measure indicates that the model achieves both low false positives and low
false negatives. In cases where class prediction is based upon class probability, logloss measure is
preferred over both Accuracy and F1 Measure. This is because logloss is a probabilistic measure
and it penalizes the predictions which are less certain based upon the prediction probability. It is
formulated as below.

91
In the above formula, yij is a binary indicator variable which denotes whether sample i belongs to
class j or not and pij indicates the probability of sample i belonging to class j. M is the total
number of classes and N is the total number of samples. The lower the value of logloss, the more
accurate is the model. Given the imbalanced nature of classes in the input SF crime dataset, using
F1 measure and log loss scores ensure that we don’t solely rely on accuracy which can be
misleading in such scenarios.
For all the models, we used 10-fold cross validation. The modelling results obtained for different
classification techniques on the validation set are presented in table 1. In this table, both Gradient
Boosting as well as Logistic Regression obtain the best overall values for F1 measure and
Logloss score. We chose Logistic Regression as our model of choice because of its faster runtime
than that of Gradient Boosting model. We submitted this Logistic Regression based model onto
Kaggle and got the rank 752 out of 2239 total submissions (in oct 2016).
Table 1. Model Evaluation results for SF crime category classification.
Algorithm F1 Measure Logloss
Logistic Regression 0.154 2.53041
Random Forest 0.190 16.7994
Gradient Boosting 0.151 2.53044
Naïve Bayes 0.154 2.8516
4.3. Model Optimization
We re-performed the model building task after grouping crime categories into three super classes.
It is a common practice in criminal law to classify crimes into three categories, namely,
Misdemeanors, Infractions and Felonies [15]. Infractions are the mildest of the crimes. Crimes
which are more serious are labeled as misdemeanors and the most serious of all crimes are
labeled as felonies. The constructed crime subgroups are displayed in table 2.
After the above categorization, we performed the modeling and validation again for our best
performing model, Logistic Regression and found that the F1 score increased to 0.59 and the log-
loss ratio reduced to 0.67. These results are very encouraging because even if the law
enforcement agencies can identify the patterns of locations and times which are more prone to
crime super classes, infractions, misdemeanors or felonies, it will be very useful for taking
concrete steps for planning police task forces for law enforcement and crime prevention.
5. UNSUPERVISED LEARNING METHODOLOGIES TO ANALYZE SF CRIME
DATASET
5.1. Motivation
Given the raw data of crimes in SF (latitude and longitude), we used contour or density plots to
study the spread of different crime types. Two of these plots are presented in figures 8 and 9 for
crime types larceny/theft and prostitution. This illustrates that crimes belonging to different
categories are localized to specific geographical areas in SF. We further plotted the normalized
crime counts for top 10 crimes to see some preliminary trends that how they vary with time

92
(figures 10 and 11). This spatial and temporal variation of crimes motivated us to study the
similarities between crimes of different categories to aid in the understanding of modus operandi
of these crimes over time and space. We used two graph based unsupervised learning methods
called clustering and core periphery structures to analyze a network of crimes built from the SF
crime dataset. We illustrate that in this case, core periphery structures are more suited for analysis
of dense graphs that vary over time. Next we describe our methodology used for SF crime
network construction and the clustering and core periphery results for this network.
Table 2. SF crime category Super Classes.
Crime Super Class Crime Types in this Super Class
Infraction LOITERING
Misdemeanor BAD CHECKS, BRIBERY, DISORDERLY CONDUCT,
DRUG/NARCOTIC, DRIVING UNDER THE INFLUENCE,
DRUNKENNESS, EMBEZZLEMENT, FAMILY OFFENSES,
FORGERY/COUNTERFEITING, GAMBLING, LIQUOR LAWS,
MISSING PERSON, NON-CRIMINAL, OTHER OFFENSES,
PORNOGRAPHY/OBSCENE MAT, PROSTITUTION, RECOVERED
VEHICLE, RUNAWAY, SECONDARY CODES, STOLEN PROPERTY,
SUICIDE, SUSPICIOUS OCC, TREA, TRESPASS, WARRANTS
Felony ARSON, ASSAULT, BURGLARY, EXTORTION, FRAUD,
KIDNAPPING, LARCENY/THEFT, ROBBERY, SEX OFFENSES
FORCIBLE, SEX OFFENSES NON-FORCIBLE, VANDALISM,
VEHICLE THEFT, WEAPON LAWS
Figure 8. Contour plot for spatial crime density distribution of crime type Larceny/Theft

93
Figure 9. Contour plot for spatial crime density distribution of crime type Prostitution
Figure 10. Variation of normalized crime count for top 10 crimes with years

94
Figure 11. Variation of normalized crime count for top 10 crimes with months
5.2. SF Crime Network Construction
Using the raw SF crime dataset, we constructed a graph G = (V, E) containing a vertex
corresponding to each of the 39 crime categories. The edge weights among different crime types
were constructed as the Tanimoto similarity [33] calculated between crime types using their
geographical proximity in terms of shared zip codes. Tanimoto similarity is an advanced version
of both cosine similarity as well as Jaccard coefficient. It assumes the calculation of similarity
between vectors whose attributes may or may not be binary. In case the vectors are binary, it
reduces to Jaccard coefficient. Given two crime types expressed as a vector of zip codes A and B,
the calculation of Tanimoto similarity between them is formalized as below.
We used a cutoff of 0.2 as the minimum similarity level required between any two crime
categories to establish an edge between the two. This led to the construction of a weighted graph
of crime types with weights as Tanimoto similarities. Further, one graph was constructed for each
of the years 2005, 2010 and 2015 to analyze temporal variation of crime trends. In table 3, we list
the global graph clustering coefficient [34] of each of these constructed graphs for the three
years. The global graph clustering coefficient is a ratio of the number of closed triplets to number
of all triplets in the graph. It gives a measure of density in the graph and lies between 0 and 1.
The high clustering coefficients for graphs of all years denote that all these graphs are very dense
in nature.
Table 3. Model Evaluation results for SF crime category classification.
Year for which Graph is constructed Graph clustering coefficient
2005 0.943
2010 0.962
2015 0.849

95
5.3. Clustering Results for SF Crime Network
We used a state-of-the-art graph clustering algorithm called ClusterONE [35] to find clusters in
all the three SF Crime graphs for years 2005, 2010 and 2015. ClusterONE uses a greedy growth-
based algorithm to grow clusters from seeds. We ran this algorithm with its default set of
parameters. The clustering results are shown in figure 12. In this figure, we can see that for the
years 2005 and 2010, all the nodes are being clustered into one big cluster. For the year 2015, all
crime categories, except Sex Offences Non-Forcible, Bad Checks, Prostitution and Extortion
form one big cluster. This behavior is due to the high graph density or cliquish nature of the
graph. In all these graphs, the width of the edges has been drawn proportional to their similarity
value. It is to be noted that the edge similarity in these graphs varies amongst different crime
types even when they are heavily connected in terms of topology.
Figure 12. Clustering results for the graph of SF crime types for the years 2005, 2010 and 2015 using
ClusterONE algorithm
Most of the graph clustering algorithms including ClusterONE [35] are good at identifying
clusters which have high topological density in the graph and give less importance to the edge
weight-based density. Thus, as the next step, we used a graph core periphery finding algorithm
called CP-MKNN [36] [37] which uses both topological density as well as edge weight density in
a graph to extract very dense core clusters surrounded by sparse periphery clusters. A core-
periphery algorithm based on the extension of ClusterONE clustering algorithm has been
proposed in [38], however, we decided to use the algorithm CP-MKNN as our choice of core
periphery algorithm because as demonstrated in [37], it is better than [38] in separating very
dense cores from their sparser surroundings. Further, there is an advantage in using core
periphery structures over simple clusters to study graphs which vary over time because we can
study how different nodes move from cores to peripheries and vice versa. This movement of data

96
can really help us in studying change dynamics, which in the current case of crime analysis can
help in understanding how crime patterns evolve over time.
5.4. Core Periphery Structures for SF Crime Network
In figure 13, we demonstrate the core periphery structures obtained by CP-MKNN [36] [37] for
the SF crime graphs for the years 2005, 2010 and 2015. We set input parameter K = 5.
Figure 13. Core Periphery structure results for the graph of SF crime types for the years 2005, 2010 and
2015 using CP-MKNN algorithm
Crimes like Vehicle Theft, Robbery and Fraud are always in the core cluster for the three years
2005, 2010 and 2015. Similarly, crimes like gambling and bribery are always in the periphery for
all the three years. There are some crime types which keep moving between the core and
periphery clusters as the years pass by. For example, the crime type Family Offences is a
peripheral crime in 2005 and 2015 and became a core crime in the year 2010. The crime type
Disorderly Conduct was in periphery in the year 2005 but a part of core cluster for the years 2010
and 2015.
This type of rich information about how crime patterns of different types of crimes change over
time with respect to each other can be highly beneficial for the law enforcement agencies to plan
their task forces for tackling anticipated crimes in different locations. Further, this information
can also help in improving awareness among the general public about the changing nature of
crimes in their neighborhoods and help them be safe.
6. CONCLUSION
In this paper, we provide an extensive survey of trends of both supervised as well as unsupervised
machine learning techniques that exist for the analysis of crimes. We use a dataset of crimes in

97
San Francisco to build supervised and unsupervised learning strategies to extract useful patterns
and trends from it.
Specifically, for supervised algorithms, we proposed a novel way of feature extraction and built
models using Logistic Regression, Gradient Boosting, Naive Bayes and Random Forest. Based
upon model evaluation results using F1 measure and Logloss score, Gradient Boosting and
Logistic Regression performed the best amongst all models. Moreover, Logistic Regression ran
much faster than the Gradient Boosting model. We further optimized this model by grouping
crime categories into super classes: “Infraction”, “Misdemeanor” and “Felony”. Our modelling
and optimization workflow can be used as a general framework to predict crime categories/ super
classes in crime incident reports from various counties. This will be highly beneficial to the law
enforcement agencies to do a risk analysis of different locations and times which are more prone
to crimes of different types.
In terms of unsupervised learning methodologies, we illustrated the use of core periphery
structures to analyze patterns in crimes that vary over time and space. We provided examples of
how certain crimes move between cores and peripheries over time while some crime types
remain stable over years. Such a behavioral study of crimes can be extremely beneficial to the
police to plan their task forces based upon the changing nature of crimes.
In a nutshell, we provide a demonstration of both supervised as well unsupervised strategies for
the analysis of crimes in San Francisco. Our modelling and optimization workflow can be used as
a general framework to predict crime categories and analyze crime behaviors using crime
incident reports from various counties all over the world. Our approaches can be immensely
helpful in planning measures for crime prevention and raising awareness among the citizens of a
county.
As a future research direction, for the unsupervised approach, we could calculate an advanced
version of distance between two crime types taking into account the actual distance in miles
between the zip codes where crimes occur. This will improve upon the accuracy of the patterns
found using clustering and core periphery structures. This will further help us in finetuning the
clusters of crime which in turn can be used as super classes in the supervised learning
optimization approach to improve its accuracy and applicability.
REFERENCES
[1] H. Adel, M. Salheen, and R. Mahmoud, “Crime in relation to urban design. Case study: the greater
Cairo region,” Ain Shams Eng. J., vol. 7, no. 3, pp. 925-938, 2016.
[2] S. Prabakaran, and S. Mitra, “Survey of analysis of crime detection techniques using data mining and
machine learning,” J. Physics: Conference Series, vol. 1000, no. 1, 2018.
[3] Q. Zhang, P. Yuan, Q. Zhou, and Z. Yang, “Mixed spatial-temporal characteristics based crime hot
spots prediction,” in IEEE 20th Intl. Conf. on Comput. Supported Cooperative Work in Des.
(CSCWD), 2016, pp. 97-101.
[4] E. Cesario, C. Catlett, and D. Talia, “Forecasting crimes using autoregressive models,” in IEEE 14th
Intl Conf on Dependable, Autonomic and Secure Computing, 14th Intl Conf on Pervasive Intelligence
and Computing, 2nd Intl Conf on Big Data Intelligence and Computing and Cyber Science and
Technology Congress (DASC/PiCom/DataCom/CyberSciTech, 2016, pp.795-802.
[5] S. Kim, P. Joshi, P.S. Kalsi, and P. Taheri, “Crime analysis through machine learning,” in IEEE 9th
Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON),
2018, pp. 415-420.
[6] N.O. Edoka, “Crime Incidents Classification Using Supervised Machine Learning Techniques:
Chicago,” PhD. Diss., Dublin, National College of Ireland, 2020.

98
[7] Y. Lamari, B. Freskura, A. Abdessamad, S. Eichberg, and S. de Bonviller, “Predicting Spatial Crime
Occurrences through an Efficient Ensemble-Learning Model.” ISPRS Int. Journal of Geo-Info, vol. 9,
no. 11, 2020.
[8] P. Cichosz, “Urban Crime Risk Prediction Using Point of Interest Data,” ISPRS Int. J. Geo-Info., vol.
9, no. 7, 2020.
[9] J. Friedman, T. Hastie, and R. Tibshirani, The elements of statistical learning, vol. 1, no. 10, New
York: Springer series in statistics, 2001.
[10] K. Ramasubramanian, and A. Singh, "Machine Learning Model Evaluation." Machine Learning
Using R, Apress, Berkeley, CA, pp. 425-464, 2017.
[11] M. Vaquero Barnadas, “Machine learning applied to Crime Prediction,” BS thesis, Universitat
Politcnica de Catalunya, 2016.
[12] A. Chandrasekar, A. Sunder Raj, and P. Kumar. “Crime prediction and classification in San Francisco
City,” URL http://cs229. stanford.edu/proj2015/228_report.pdf, 2015.
[13] I. Pradhan, “Exploratory Data Analysis and Crime Prediction in San Francisco,” MS Project, San
Jose State University, USA, 2018.
[14] S. T. Ang, W. Wang, and S. Chyou, “San Francisco crime classification,” UCSD Project Report URL
https://cseweb.ucsd.edu/classes/wi15/cse255-a/reports/fa15/037.pdf, 2015.
[15] Y. Abouelnaga, “San Francisco crime classification,” arXiv preprint arXiv:1607.03626, 2016.
[16] X. Wu, “An informative and predictive analysis of the San Francisco police department crime data,”
Ph.D. Diss., UCLA, 2016.
[17] “Classifications of Crime”, FindLaw, URL https://criminal.findlaw.com/criminal-law-
basics/classifications-of-crimes.html
[18] A. Sharma, and S. Sharma, “An intelligent analysis of web Crime Data using Data Mining,” Int. J.
Eng. and Innov. Tech. (IJEIT), vol. 2, no. 3, 2012.
[19] S. Kumar, and D. Toshniwal, “A Data Mining framework to analyze Road Accident Data,” J. Big
Data, vol. 2, no. 1, pp. 26, 2015.
[20] A. Joshi, A. Sai Sabitha, and T. Choudhury, “Crime analysis using K- means clustering,” in 3rd Int.
Conf. on Comp. Intell. and Net. (CINE), pp. 33-39, 2017.
[21] B. Sivanagaleela, and S. Rajesh, “Crime Analysis and Prediction Using Fuzzy C-Means Algorithm,”
in 3rd Int. Conf. on Trends in Electronics and Informatics (ICOEI), 2019, pp. 595-599.
[22] C. Morselli, Inside criminal networks, vol. 8, New York, Springer, 2009.
[23] J. Xu, and H. Chen, “The topology of dark networks,” Communications of the ACM, vol. 51, no. 10,
pp. 58-65, 2008.
[24] F. Calderoni, “The structure of Drug Trafficking mafias: the Ndrangheta and cocaine,” Crime, law
and social change, vol. 58, no. 3, pp. 321-349, 2012.
[25] T. Divik, J. K. Dijkstra, and T. A. B. Snijders, “Structure, multiplexity, and centrality in a corruption
network: the Czech Rath affair,” Trends in Organized Crime, vol. 22, no. 3, pp. 274-297, 2019.
[26] A. Kantamneni, “Identifying Communities as Core-Periphery Structures in Evolving Networks,” MS
Thesis University of Cincinnati, 2016.
[27] G. Berlusconi, F. Calderoni, N. Parolini, N., M. Verani, and C. Piccardi, “Link prediction in criminal
networks: A tool for criminal intelligence analysis,” PloS one, vol. 11, no. 4, 2016.
[28] E. Budur, S. Lee, and V. S. Kong, “Structural analysis of Criminal Network and predicting hidden
links using Machine Learning,” URL arXiv:1507.05739, 2015.
[29] “San Francisco Crime Data and Arrest Information”, Civic Hub, URL
https://www.civichub.us/ca/san-francisco/gov/police-department/crime-data.
[30] “San Francisco Crime Classification”, Kaggle, URL https://www.kaggle.com/c/sf-crime.
[31] “Police Department Incident Reports”, DataSF, URL https://data.sfgov. org/Public-Safety/Police-
Department-Incident-Reports-Historical-2003/tmnf-yvry/.
[32] N. Van Gheem, “Pyzipcode 1.0. 2016," URL https://pypi.org/project/pyzipcode/, 2016.
[33] A.A. Goshtasby, "Similarity and dissimilarity measures," Image registration. Springer, London, pp. 7-
66, 2012.
[34] A. Barrat, M. Barthelemy, R. Pastor-Satorras and A. Vespignani, "The architecture of complex
weighted networks," in Proc. Nat. Acad. Sci., 2004, vol. 101, no. 11, pp. 3747-3752.
[35] T. Nepusz, H. Yu, and A. Paccanaro, “Detecting overlapping protein complexes in Protein-Protein
Interaction Networks,” Nature Methods, vol. 9, pp. 471-472, 2012.
[36] D. Sardana, “Analysis of Meso-scale Structures in Weighted Graphs,” PhD diss., University of
Cincinnati, 2017.

99
[37] D. Sardana and R. Bhatnagar, “Graph Algorithm to find Core Periphery structures using Mutual K-
Nearest Neighbors” Int. J. Art. Intell. (IJAIA): in press, 2021.
[38] D. Sardana and R. Bhatnagar, “Core Periphery structures in weighted graphs using Greedy growth,”
in IEEE/WIC/ACM International Conference on Web Intelligence, 2016, pp. 1-8.
AUTHORS
Divya Sardana: Dr. Divya Sardana is a Senior Data Scientist at Teradata Corp.,
Santa Clara, CA, USA. She has a PhD in Computer Science from the University of
Cincinnati, OH US. Her research targets development of scalable machine learning
and graph algorithms for the analysis of complex datasets in interdisciplinary
domains of data science such as Bioinformatics, Sociology and Smart
Manufacturing.
Shruti Marwaha: Dr Shruti Marwaha is a Research Engineer in the Department of
Cardiovascular Medicine. She received her PhD in Systems Biology and
Physiology from University of Cincinnati in 2015 and has been working as a
Bioinformatician with Stanford Center for Undiagnosed Diseases since 2016. Her
primary role involves analysis and management of genomics and immunology
data. She focuses on benchmarking variant callers, optimizing tools for variant
prioritization and analysis of single cell immune data.
Raj Bhatnagar: Dr. Raj Bhatnagar is a Professor at the department of EECS at
University of Cincinnati, Cincinnati, OH, USA. His research interests encompass
developing algorithms for data mining and pattern recognition problems in various
domains including Bioinformatics, Geographic Information Systems,
Manufacturing, and Business.

SUPERVISED AND UNSUPERVISED MACHINE LEARNING METHODOLOGIES FOR CRIME PATTERN ANALYSIS

More Related Content

What's hot (13)

Similar to SUPERVISED AND UNSUPERVISED MACHINE LEARNING METHODOLOGIES FOR CRIME PATTERN ANALYSIS (20)

Recently uploaded (20)

SUPERVISED AND UNSUPERVISED MACHINE LEARNING METHODOLOGIES FOR CRIME PATTERN ANALYSIS