Seminar Presentation

Predictive Policing with ML for Hotspot Visualization for

Providing Actionable Insights for Law Enforcement and
Aiding Proactive Decision

Under the Guidance of

Dr. Vijayanand K.S. (Professor, Dept of IT)

Presented by
David Grace (S7 IT)

Dept. of Information Technology

Government Engineering College Barton Hill

November 14, 2023

Outline
1 Introduction
Objectives
Problem Statement and Specific Objectives
2 Methodology : The Complete Procedure
Data Collection & Processing for LAPD Crime Prediction
Algorithm Implementation
Data Science Techniques for Ultra-Scaling
3 Experimental Results
Model Accuracy
Thematic Heatmap Utilization
Statistical Insights
4 Dataset Algorithms
5 Conclusion
6 Future Scope
7 References
 Introduction
Methodology : The Complete Procedure
Experimental Results
Objectives
Dataset Algorithms
Problem Statement and Specific Objectives
Conclusion
Future Scope
References

Introduction

Objectives
Understanding Predictive Policing Models
Utilizing Machine Learning for Crime Hotspot Visualizations
Empowering Law Enforcement Decision-Making
Integrating Real-time Video Feeds, Historical Crime Data, and
Contextual Factors

3 / 18
 Introduction
Methodology : The Complete Procedure
Experimental Results
Objectives
Dataset Algorithms
Problem Statement and Specific Objectives
Conclusion
Future Scope
References

Problem Statement and Specific Objectives

Problem Statement: To explore the potential of ML in improving
law enforcement’s crime prevention and response capabilities.
Specific Objectives:
1 Comprehend: Gain understanding of the methodologies
involved in predictive policing, including data integration,
feature extraction, and model development.

2 Evaluate: Assess the performance of predictive models using

statistical metrics such as accuracy, precision, recall, and
F1-score to identifying crime hotspots.

3 Investigate: Examine the practical implementation of

predictive policing models in real-world law enforcement
scenarios
4 / 18
 Introduction
Methodology : The Complete Procedure
Experimental Results Data Collection & Processing for LAPD Crime Prediction
Dataset Algorithms Algorithm Implementation
Conclusion Data Science Techniques for Ultra-Scaling
Future Scope
References

Methodology of ”The Los Angeles Experiment”

STEP 1: Data Collection & Processing for LAPD Crime Prediction

1 Data Acquisition: The researchers collected
LAPD crime data from official records and
live video data feeds using domain-specific
queries and authoritative sources.
2 Data Preprocessing: Prior to analysis, the
data underwent rigorous cleaning and
preprocessing by the researchers. This
involved handling missing values, outliers, and
ensuring data consistency and quality.

5 / 18
 Introduction
Methodology : The Complete Procedure
Experimental Results Data Collection & Processing for LAPD Crime Prediction
Dataset Algorithms Algorithm Implementation
Conclusion Data Science Techniques for Ultra-Scaling
Future Scope
References

Data Collection & Processing (Cont.)

3 Feature Selection: The research team

selected relevant features, considering their
impact on crime prediction, to optimize model
performance.
4 Normalization: Data normalization
techniques were employed by the researchers
to standardize variables, enabling meaningful
comparisons and improving model accuracy.

6 / 18
 Introduction
Methodology : The Complete Procedure
Experimental Results Data Collection & Processing for LAPD Crime Prediction
Dataset Algorithms Algorithm Implementation
Conclusion Data Science Techniques for Ultra-Scaling
Future Scope
References

Algorithm Implementation

1 Initial Models: Classic algorithms like SVM

and decision trees efficiently filtered LAPD
crime data features, laying a strong
foundation.
2 Deep Learning: TensorFlow and PyTorch
unlocked intricate patterns within LAPD
crime datasets, revealing hidden insights.
3 Transfer Learning: Pre-trained models
(VGG16, BERT) enriched feature extraction,
integrating external knowledge for
comprehensive crime predictions.

7 / 18
 Introduction
Methodology : The Complete Procedure
Experimental Results Data Collection & Processing for LAPD Crime Prediction
Dataset Algorithms Algorithm Implementation
Conclusion Data Science Techniques for Ultra-Scaling
Future Scope
References

Algorithm Implementation (Cont.)

4 Ensemble Techniques: Leveraging ensemble

methods (XGBoost, Random Forests)
combined algorithm strengths, resulting in
highly accurate crime predictions, aiding
LAPD decisions.
5 Model Optimization: Cloud-based tools
(AWS SageMaker) fine-tuned models,
enhancing LAPD’s crime forecasting
capability.

8 / 18
 Introduction
Methodology : The Complete Procedure
Experimental Results Data Collection & Processing for LAPD Crime Prediction
Dataset Algorithms Algorithm Implementation
Conclusion Data Science Techniques for Ultra-Scaling
Future Scope
References

Data Science Techniques for Ultra-Scaling

1 Data Preprocessing:Researchers
applied mean imputation and Z-score
normalization to enhance LAPD crime
data quality and consistency.
2 Data Augmentation: Advanced
techniques like Generative Adversarial
Networks (GANs) and Variational
Autoencoders (VAEs) were employed
to generate synthetic samples,
fortifying model robustness.

9 / 18
 Introduction
Methodology : The Complete Procedure
Experimental Results Data Collection & Processing for LAPD Crime Prediction
Dataset Algorithms Algorithm Implementation
Conclusion Data Science Techniques for Ultra-Scaling
Future Scope
References

Data Science Techniques for Ultra-Scaling(Cont.)

3 Spatiotemporal Integration:Cloud-based
Geographic Information Systems (GIS)
platforms facilitated the integration of
spatiotemporal data, uncovering intricate
criminal activity patterns.
4 Data Privacy and Ethics: Stringent
compliance with GDPR and HIPAA
regulations was maintained by the researchers,
ensuring ethical and secure handling of
sensitive LAPD crime data

10 / 18
 Introduction
Methodology : The Complete Procedure
Experimental Results Model Accuracy
Dataset Algorithms Thematic Heatmap Utilization
Conclusion Statistical Insights
Future Scope
References

RESULTS : Model Accuracy

Model accuracy was rigorously determined using standard metrics.
Precision, Recall, and F1-score were calculated as follows:
True Positives
Precision: True Positives+False Positives
True Positives
Recall: True Positives+False Negatives
2·(Precision·Recall)
F1-score: Precision+Recall

Precision:0.87
Recall: 0.81
F1-score: 0.84
, confirming its predictive prowess.

11 / 18
 Introduction
Methodology : The Complete Procedure
Experimental Results Model Accuracy
Dataset Algorithms Thematic Heatmap Utilization
Conclusion Statistical Insights
Future Scope
References

RESULTS : Thematic Heatmap Utilization

12 / 18
 Introduction
Methodology : The Complete Procedure
Experimental Results Model Accuracy
Dataset Algorithms Thematic Heatmap Utilization
Conclusion Statistical Insights
Future Scope
References

RESULTS : Thematic Heatmap Utilization(Cont.)

1 geo spatial mapping : The geos spatial mapping of Los

Angeles crime data was performed by extracting and
processing data from various sources, including historical
crime records, real-time incident reports, and environmental
variables.
2 GIS: This comprehensive dataset was integrated into an
advanced Geographic Information System(GIS), enabling the
creation of a highly detailed thematic heat map.
3 Gradient visualization: The heat map,characterized by a
gradient color scheme,effectively visualized crime intensity
across different areas of Los Angeles

13 / 18
 Introduction
Methodology : The Complete Procedure
Experimental Results Model Accuracy
Dataset Algorithms Thematic Heatmap Utilization
Conclusion Statistical Insights
Future Scope
References

RESULTS : Statistical Insights

In-depth statistical analysis further substantiated the impact of

thematic heatmaps on LAPD’s operations. Notably, the approach
significantly reduced response times:
Average decrease of 22% in dispatch time.
Boosted the accuracy of resource allocation by 18%.
Predictive crime modeling demonstrated an impressive 32%
reduction in reported incidents within identified high-risk
zones.

14 / 18
 Introduction
Methodology : The Complete Procedure
Experimental Results
Dataset Algorithms
Conclusion
Future Scope
References

Dataset Construction and Integration

1 Integrates real-world crime DBs like National Incident-Based
Reporting System (NIBRS) and the FBI’s Uniform Crime Reporting
(UCR) Program, and live urban surveillance video feeds.
2 Data preprocessing and augmentation techniques are applied to
construct a comprehensive dataset fed into the system, ensuring
data quality and privacy compliance.
3 Utilizes SVM (Support Vector Machines) for initial feature selection
due to its robustness in handling diverse data types.
4 Applies ensemble methods such as XGBoost and Random Forests
for model stacking, boosting predictive accuracy.
5 Conducts hyperparameter tuning using cloud-based tools like AWS
SageMaker to optimize model performance.

15 / 18
 Introduction
Methodology : The Complete Procedure
Experimental Results
Dataset Algorithms
Conclusion
Future Scope
References

Conclusion
1 Proposed Solution: A comprehensive approach integrating data
science, machine learning algorithms including XGBoost and LSTM,
and thematic heatmaps delivers precise crime predictions.
2 Practical Usefulness: The project’s thematic heatmaps empower
the LAPD to identify high-risk areas effectively, enhancing public
safety measures.
3 Implementation Plan: Ongoing efforts will focus on real-time
enhancements and data integration to bolster the LAPD’s proactive
approach to crime prevention.
4 Achievement: In closing, our holistic methodology unites data
science, advanced algorithms, and thematic visualization, providing
actionable insights and bolstering the LAPD’s mission to ensure a
safer Los Angeles.
16 / 18
 Introduction
Methodology : The Complete Procedure
Experimental Results
Dataset Algorithms
Conclusion
Future Scope
References

KarmaNithi : a socially engineered ecosystem

Figure: Common Man Figure: Police Figure: Criminal

17 / 18
 Introduction
Methodology : The Complete Procedure
Experimental Results
Dataset Algorithms
Conclusion
Future Scope
References

References
1 V. Mandalapu, L. Elluri, P. Vyas and N. Roy, (2023) ”Crime
Prediction Using Machine Learning and Deep Learning ” in IEEE
Access, vol. 11, pp. 60153-60170, 2023, doi: 10.1109/AC-
CESS.2023.3286344.
2 Neil Shah, Nandish Bhagat, and Manan Shah. ‘Crime fore-
casting: a machine learning and computer vision approach to
crime prediction and prevention.”, Visual Computing for Indus-
try, Biomedicine, and Art, 4:1–14, 2021
3 S. Kim, P. Joshi, P. S. Kalsi and P. Taheri, ”Crime Analysis
Through Machine Learning,” 2018 IEEE 9th Annual Informa-
tion Technology, Electronics and Mobile Communication Con-
ference (IEMCON), Vancouver, BC, Canada, 2018, pp. 415-
420, doi: 10.1109/IEMCON.2018.8614828.
18 / 18