crime type and occurrence prediction using machine learning

Abstract:

This project aims to develop a machine learning model for crime type
and occurrence prediction based on historical data. The process
involves collecting diverse datasets encompassing crime reports,
demographics, weather conditions, and socioeconomic factors. After
thorough preprocessing and feature engineering, exploratory data
analysis provides insights into underlying patterns and relationships.
Feature selection is performed to identify the most influential variables,
and a suitable machine learning algorithm, such as Random Forests, is
selected for modeling. The model is trained on a split dataset, and
hyperparameters are tuned for optimal performance. Evaluation metrics,
including accuracy, precision, recall, and F1-score, are utilized to assess
the model's effectiveness. Deployment of the model for real-time
predictions is followed by continuous monitoring and improvement,
ensuring adaptability to changing circumstances. Ethical considerations,
transparency, and collaboration with domain experts are integral to
addressing potential biases and ensuring the model's fairness and
usefulness in real-world scenarios
Introduction
Crime prediction has been a longstanding challenge for law enforcement
agencies seeking to enhance proactive strategies and community safety.
With the advent of machine learning techniques, there is a growing
opportunity to leverage historical data to predict the types and
occurrences of crimes. This project delves into the development of a
machine learning model designed to forecast crime patterns based on a
comprehensive dataset that includes crime reports, demographics,
weather conditions, and socioeconomic indicators.

Understanding and predicting crime is a multifaceted task that involves

the intricate interplay of various factors. Machine learning offers a
promising avenue to discern hidden patterns within vast datasets,
providing law enforcement agencies with valuable insights to allocate
resources efficiently and implement preventative measures. By
analyzing historical crime data, this project aims to create a predictive
model capable of anticipating not only the occurrence of crimes but also
their specific types.

The journey begins with the collection of diverse and representative

datasets, ensuring a comprehensive understanding of the factors
influencing criminal activities. Subsequent steps involve meticulous data
preprocessing, feature engineering, and exploratory data analysis to
uncover underlying trends. The selection of relevant features is a critical
aspect, determining the model's ability to generalize to new scenarios.

A key consideration in this endeavor is the ethical implications

associated with crime prediction. It is imperative to ensure that the model
is fair, unbiased, and transparent, addressing concerns related to privacy
and discrimination. Collaboration with domain experts and stakeholders
is vital throughout the development process to align the model with the
needs of the community and law enforcement.

As the project progresses, the chosen machine learning algorithm

undergoes training and evaluation, with performance metrics guiding the
refinement of the model. Deployment for real-time predictions is followed
by continuous monitoring and improvement to adapt to evolving crime
patterns. The goal is to create a reliable and effective tool that aids law
enforcement agencies in their mission to enhance public safety through
proactive crime prediction.

Object

The primary objective of this project is to develop a robust and accurate

machine learning model for crime type and occurrence prediction. The
specific goals include:

1. Data Collection and Integration:

 Gather diverse datasets encompassing historical crime reports,
demographic information, weather conditions, and socioeconomic
factors. Ensure data integrity, relevance, and representation.

2. Data Preprocessing and Cleaning:

Perform comprehensive data cleaning and preprocessing to handle
missing values, outliers, and inconsistencies. Convert categorical
variables into numerical formats and standardize data for uniformity.

3. Feature Engineering:
Explore and engineer features that capture temporal, spatial, and
contextual aspects of crime occurrences. Extract relevant information
from timestamps, geographical coordinates, and other available data to
enhance predictive capabilities.

4. Exploratory Data Analysis (EDA)

Conduct exploratory data analysis to gain insights into patterns,
correlations, and anomalies within the dataset. Use visualizations and
statistical techniques to inform feature selection and model design.

5. Feature Selection:
- Identify and select the most influential features for crime prediction.
Employ techniques such as correlation analysis, recursive feature
elimination, or feature importance scores from machine learning models.

6. Model Selection:
Choose a suitable machine learning algorithm (e.g., Random Forests,
Support Vector Machines) based on the nature of the crime prediction
task. Consider ensemble methods for robustness and scalability.

7. Model Training and Hyperparameter Tuning:

Split the dataset into training and testing sets. Train the selected
model on the training set and fine-tune hyperparameters to optimize
performance. Ensure the model's ability to generalize to unseen data.
8. Evaluation Metrics:
Employ appropriate evaluation metrics, such as accuracy, precision,
recall, and F1-score, to assess the model's performance. Iterate on the
model based on evaluation results.

9. Deployment:
Deploy the trained model for real-time crime predictions. Ensure
seamless integration with existing systems and evaluate the model's
effectiveness in practical scenarios.

10. Continuous Monitoring and Improvement:

Implement a system for continuous monitoring of the model's
performance. Regularly update the model with new data and refine it to
adapt to evolving crime patterns. Address any emerging issues promptly.

11. Ethical Considerations:

Prioritize ethical considerations throughout the project, ensuring
fairness, transparency, and accountability. Mitigate potential biases and
collaborate with domain experts to address ethical concerns associated
with crime prediction.

By accomplishing these objectives, the project aims to contribute to the

development of a reliable and ethical crime prediction tool that enhances
proactive law enforcement strategies and community safety.

List review

1. Model Performance Evaluation:

Accuracy and Precision:

Evaluate the accuracy and precision of various machine learning models
in predicting crime types and occurrences. Examine how well these
models perform on diverse datasets and their ability to handle
imbalanced classes.

Comparison of Algorithms:

Compare the performance of different machine learning algorithms, such

as Random Forests, Support Vector Machines, and Neural Networks, in
terms of prediction accuracy and computational efficiency.
Evaluation Metrics:

Discuss the choice of evaluation metrics (e.g., F1-score, ROC-AUC)

and their significance in assessing the effectiveness of crime prediction
models. Analyze the trade-offs involved in selecting specific metrics
based on the goals of the prediction task.

2. Ethical Considerations and Bias Mitigation:

Fairness and Bias:

Examine the ethical considerations associated with crime prediction

models, addressing issues of fairness, transparency, and potential
biases. Evaluate how models perform across different demographic
groups and socioeconomic contexts.

Privacy Concerns:

Discuss the privacy implications of deploying crime prediction models

and assess measures taken to protect individual privacy rights. Explore
methods for anonymizing data and ensuring compliance with privacy
regulations.

Community Impact:
Investigate the impact of deploying crime prediction models on
communities. Assess whether these models lead to positive outcomes in
terms of public safety and community trust, or if there are unintended
consequences that need addressing
.
3. Real-World Deployment and Practicality
Scalability and Integration:

Evaluate the scalability of crime prediction models for deployment in

real-world law enforcement systems. Discuss the challenges and
solutions for integrating these models into existing workflows and
decision-making processes.

User-Friendliness:

Assess the user-friendliness of crime prediction tools for law

enforcement personnel. Explore the ease of interpretation and usability
of these tools in supporting proactive strategies.

Continuous Improvement and Adaptability:

Investigate the mechanisms in place for continuous monitoring and

improvement of crime prediction models. Discuss how models adapt to
evolving crime patterns and whether they effectively learn from new data
to enhance predictive capabilities.

Existing System

In pre-work, the dataset obtained from the open source are first pre-
processed to remove the duplicated values and features.
  Decision tree has been used in the factor of finding crime patterns
and also extracting the features from large amount of data is
inclusive. It provides a primary structure for further classification
process.
 The classified crime patterns are feature extracted using Deep
Neural network. Based on the prediction, the performance is
calculated for both trained and test values.
 The crime prediction helps in forecasting the future happening of
any type Proceedings of the International Conference on Artificial
Intelligence and Smart Systems of criminal activities and help the
officials to resolve them at the earliest.
Drawbacks
1.The pre-existing works account for low accuracy since the classifier
uses a categorical values which produces a biased outcome for the
nominal attributes with greater value.

2.The classification techniques does not suited for regions with

inappropriate data and real valued attributes.

3. The value of the classifier must be tuned and hence there is a need of
assigning an optimal value

1. Overview:
Name: PredPol (Predictive Policing)
 Description: PredPol is a predictive policing system that employs
machine learning algorithms to forecast areas where specific types of
crimes are more likely to occur. The system analyzes historical crime
data, including the type, location, and time of incidents, to identify
patterns and trends.

2. Functionality:

Data Utilization: PredPol uses historical crime data to identify

hotspots and predict future criminal activities. It considers factors such
as time of day, day of the week, and geographic location to generate
predictions.

Algorithm: The system typically uses machine learning algorithms,

often based on a mathematical model known as the "self-exciting point
process," to predict the likelihood of crimes occurring in specific areas.

3. Key Features

Targeted Deployment: PredPol aims to guide law enforcement in

deploying resources more efficiently by concentrating efforts in predicted
high-risk areas.

Crime Types:The system is designed to predict various types of

crimes, including burglary, robbery, and vehicle theft.

Continuous Learning: PredPol continually learns from new crime

data, adapting its predictions over time to reflect evolving patterns.

4. Challenges
 Ethical Concerns: Predictive policing systems, including PredPol, have
faced criticism for potential biases in historical data and concerns about
exacerbating existing inequalities.

Transparency: There have been discussions about the need for

transparency in the algorithms used by predictive policing systems to
ensure accountability and address public concerns.

5. Deployment and Impact:

Deployment PredPol has been deployed in various cities, including Los

Angeles, Atlanta, and London, as part of efforts to enhance proactive
policing strategies.

Impact: Some studies suggest that PredPol has led to reductions in

certain types of crimes in specific areas. However, the impact and
effectiveness of predictive policing remain topics of ongoing research
and debate.

Proposed System: Crime Prediction and Prevention Platform

1. . Unique Data Sources: Do you plan to utilize unconventional data

sources beyond the usual police reports, emergency calls, and social
media? Perhaps environmental sensors, traffic flow data, or even
sentiment analysis from local news and online forums?

2. Innovative Feature Engineering: How will you extract and represent

features from the data? Imagine a more nuanced approach than typical
time, location, and weather data. Perhaps analyzing spatial
patterns, seasonal trends, or even correlations between crime types and
specific keywords in social media discussions
3. Choice of Machine Learning Algorithms: Are you considering any novel
algorithms or model architectures? Maybe exploring explainable AI
models like decision trees for greater transparency, or leveraging deep
learning for complex spatiotemporal analysis

4. Prediction and Visualization Techniques: How will you present predicted

crime types and probabilities? Think beyond standard heatmaps. Could
you utilize interactive dashboards, 3D crime risk maps, or even natural
language summaries to convey insights to law enforcement?

5. Addressing Ethical Considerations: How will you ensure your system is

transparent, unbiased, and used responsibly? Consider implementing
techniques like fairness-aware algorithms, data anonymization, and clear
communication channels with stakeholders.

Advantages

1. The proposed algorithm is well suited for the crime pattern detection since
most of the featured attributes depends on the time and location.

2. It also overcomes the problem of analysing independent effect of the

attributes.

3. The initialization of optimal value is not required since it accounts for real
valued, nominal value and also concern the region with insufficient
information.

4. The accuracy has been relatively high when compared to other machine
learning prediction model

Data Integration and Preprocessing Module:

Objective: Collect, integrate, and preprocess diverse datasets to create
a comprehensive input for the machine learning model.
Components:
 Data obtained from the open source must be first pre-processed in
order to overcome unnecessary violations.
 The dataset has been chosen for Denver city with enormous
amount of crime data over six years.
  The machine learning technique filter and wrapper is implied to
find the missing integral in specified attribute values.

 Data cleaning play a vital in training a prediction model and also in

the performance of the commenced process. Filtering the instance
and removal of irrelevant context from datasets are done.

 The filtering methods contributes in measuring the significance of

the features.

 The correlation with the dependent values is considered in the

feature selection.
 The wrapper method imposed is used in measuring how useful is
the feature subset by training a prediction model on it actually. The
data after pre-processing is split into test and trained attributes

 Data Collection: Interface with various sources, including crime

reports, demographic databases, weather APIs, and
socioeconomic indicators.

 Data Cleaning and Preprocessing: Implement algorithms for

handling missing values, outliers, and standardizing data formats.

Feature Engineering and Exploration Module:

Objective: Extract meaningful features from the integrated dataset and

explore patterns through data visualization and analysis.
Components:

Mapping The crime features such as crime type, the date on which the
crime has been occurred including the time of occurrence are first
segregated
 It is then mapped to an integer for easy labelling.
 The labelled details are further analysed and used are used in
graph plotting.]

 Python is chosen as programming language in implementing the

proposed work sin \
  it is well suited for machine learning process.

 The package matplotlib is imported in order to plot the graph to

show the occurrence of the criminal activities.

 The crime which occurred the most can be plotted in the graph
which contributes for further prediction process

 Feature Engineering: Develop algorithms to extract temporal,

spatial, and contextual features that enhance the model's
predictive capabilities
 .
 Exploratory Data Analysis (EDA): Utilize statistical methods and
visualizations to gain insights into patterns, correlations, and
anomalies in the dataset.

Machine Learning Model Development Module:

Objective: Select, train, and optimize machine learning models for

accurate crime type and occurrence predictions.
Components:
 The reason behind the application of Naïve Bayes is that crime
prediction usually concerns with the temporal and spatial data.

 The independent effect among the attribute values is first analysed

since the selected crime attributes possess an independent effect
upon them.

 They are used in creating a model by providing a training using

crime data that are related to robbery, burglary, murder, sexual
abusing, armed robbery, chain snatching, gang rape and highway
robbery.
Some of the extended techniques of Naïve Bayes has been implied.

1. Gaussian Naïve Bayes is related to real valued attribute selection. It

is otherwise stated as normal distribution that is done by calculating the
standard deviation and mean from the trained data.
2. Multi-nominal Naïve Bayes is applied for multiple classifier that
corresponds to the categorical features in the trained value.

3.Bernouille Naïve Bayes is used for the working of independent feature

effects of the selected attributes for crime prediction

 Model Selection: Choose an appropriate algorithm (e.g., Random

Forests) based on evaluation criteria and model requirements.

 Training and Hyperparameter Tuning: Split the dataset, train the

model, and fine-tune hyperparameters for optimal performance
.
 Evaluation Metrics: Implement metrics such as accuracy,
precision, recall, and F1-score to assess model performance.
Ethical Considerations and Bias Mitigation Module:

Objective: Address ethical concerns associated with crime prediction,

ensuring fairness, transparency, and the mitigation of biases.
Components:
 Fairness Assessment: Implement algorithms to assess and
mitigate biases in the model's predictions across different
demographic groups.
 Privacy Measures: Develop protocols for anonymizing sensitive
data and complying with privacy regulations.
 Community Impact Analysis: Evaluate the potential impact of the
system on communities, emphasizing positive outcomes and
mitigating unintended consequences.
Real-World Deployment and Continuous Improvement Module:

Objective: Deploy the model in real-world scenarios, monitor its

performance, and ensure continuous improvement.
Components:
 Scalability and Integration: Develop scalable solutions for
integrating the prediction model into existing law enforcement
systems.
 User-Friendly Interface: Design an intuitive interface for law
enforcement personnel, facilitating easy interpretation and
utilization of predictions.
 Continuous Monitoring: Implement mechanisms for ongoing
monitoring of the model's performance and adapting to evolving
crime patterns.
 Model Update and Adaptation: Establish procedures for updating
the model with new data, retraining it, and adapting to emerging
challenges.

System Analysis for Crime Prediction System Using Machine

Learning:
System Overview:
  Purpose: The primary purpose of the system is to predict crime
types and occurrences using machine learning algorithms, aiding
law enforcement agencies in proactive decision-making and
resource allocation.

 Components: The system consists of data collection modules,

preprocessing and feature engineering components, machine
learning models, evaluation metrics, deployment mechanisms, and
continuous monitoring processes

.
Data Flow:

 Input: Diverse datasets, including historical crime reports,

demographic information, weather conditions, and socioeconomic
factors.

 Processing: Data preprocessing involves cleaning, handling

missing values, and feature engineering. Machine learning models
are trained on the processed data, and hyperparameters are tuned
for optimal performanc

 Output: Crime predictions, evaluation metrics results, and insights

derived from the analysis of historical data.

Functional Components:

 Data Collection Module: Gathers and integrates diverse datasets

from crime reports, demographics, weather, and socioeconomic
indicators.

 Preprocessing and Feature Engineering: Cleans and transforms

data, extracts relevant features, and prepares the dataset for
machine learning.

 Machine Learning Model: Utilizes algorithms (e.g., Random

Forests) for training and predicting crime types and occurrences

.
  Evaluation Metrics Module: Assess the model's performance using
metrics such as accuracy, precision, recall, and F1-score.

 Deployment Module: Implements the model for real-time

predictions within law enforcement systems.

 Continuous Monitoring and Improvement: Regularly updates the

model with new data, monitors its performance, and refines it to
adapt to changing crime patterns.

Interactions:

 User Interaction: Law enforcement personnel interact with the

system through a user-friendly interface, accessing crime
predictions and relevant insights
.
 Model Interaction: The machine learning model interacts with
historical data during the training phase and with real-time data
during deployment for prediction
.
 Continuous Monitoring Interaction: Involves feedback loops where
the system continuously monitors model performance, user
feedback, and emerging crime patterns

Ethical Considerations:

 Fairness: The system must address potential biases to ensure fair

predictions across different demographic groups.

 Privacy: Measures are in place to safeguard individual privacy

rights, with anonymization techniques and compliance with privacy
regulations.

 Transparency: The system ensures transparency in how

predictions are made, providing interpretable insights for users and
stakeholders.
Scalability and Integration:

Scalability: The system is designed to scale with increasing data

volumes and user demands for predictions.
Integration: Seamlessly integrates with existing law enforcement
systems and workflows, ensuring compatibility and minimal disruption.
Usability and User Training:

User-Friendliness: The system prioritizes an intuitive interface for law

enforcement personnel, allowing easy interpretation and interaction.
Training: Users receive training on the system's functionalities,
interpretation of predictions, and understanding of ethical considerations.

Adaptability and Continuous Improvement:

Adaptability: The system is capable of adapting to evolving crime

patterns, adjusting its predictions based on new data and emerging
trends.
Continuous Improvement: Regular updates and refinements are
implemented based on continuous monitoring, user feedback, and
advancements in machine learning techniques.

Modular Description
Model Performance Evaluation:

 Accuracy and Precision: In this module, we delve into the

fundamental metrics of accuracy and precision in crime prediction
models. We assess how accurately these models identify and
classify different types of crimes. An exploration of precision is
crucial to understanding the model's ability to minimize false
positives and make reliable predictions.
  Comparison of Algorithms: This module involves a detailed
comparison of various machine learning algorithms commonly
employed in crime prediction. We evaluate their performance in
terms of accuracy, computational efficiency, and suitability for
different types of crime datasets. This comparison sheds light on
the strengths and weaknesses of each algorithm, aiding in
informed model selection.

 Evaluation Metrics: Here, we scrutinize the diverse evaluation

metrics used to assess crime prediction models. We explain the
significance of metrics like F1-score, ROC-AUC, and others in
different scenarios. The discussion includes insights into the trade-
offs associated with choosing specific evaluation metrics based on
the objectives of the crime prediction task.

Ethical Considerations and Bias Mitigation:

 Fairness and Bias: In this module, we explore the ethical

dimensions of deploying crime prediction models, emphasizing
fairness and the mitigation of biases. We analyze how models
perform across demographic and socioeconomic groups,
identifying potential biases that may impact their equitable
application in diverse communities.

 Privacy Concerns: This module investigates the privacy

implications of implementing crime prediction models. We delve
into strategies for anonymizing sensitive data and ensuring
compliance with privacy regulations. Discussion points include the
balance between effective crime prediction and safeguarding
individual privacy rights.

 Community Impact: Here, we assess the broader impact of crime

prediction models on communities. We explore whether these
models contribute positively to public safety and community trust
or if there are unintended consequences. The module considers
community engagement, feedback mechanisms, and the role of
transparency in fostering positive outcomes.
Real-World Deployment and Practicality:

 Scalability and Integration: In this module, we evaluate the

scalability of crime prediction models for practical deployment. We
discuss challenges and solutions related to integrating these
models into existing law enforcement systems. The focus is on
ensuring seamless integration and scalability to meet the demands
of real-world scenarios.

 User-Friendliness: This module assesses the user-friendliness of

crime prediction tools for law enforcement personnel. We examine
the interpretability and usability of these tools in supporting
proactive strategies. User feedback and interface design
considerations are explored to enhance the overall user
experience.

 Continuous Improvement and Adaptability: Here, we investigate

the mechanisms in place for continuous monitoring and
improvement of crime prediction models. We analyze how these
models adapt to evolving crime patterns and learn from new data.
The module emphasizes the importance of model adaptability for
sustained effectiveness in real-world applications.

Conclusion

In concluding the review of crime prediction models using machine

learning, it becomes evident that advancements in technology hold great
promise for enhancing law enforcement strategies and community
safety.

The examination of model performance underscored the importance of

choosing appropriate evaluation metrics and considering algorithmic
trade-offs. While certain models demonstrated high accuracy, their
effectiveness varied across different datasets, emphasizing the need for
a nuanced understanding of predictive capabilities.
Ethical considerations emerged as a critical aspect in deploying crime
prediction models. Addressing issues of fairness, transparency, and bias
mitigation is imperative to ensure that these tools serve communities
equitably.

Privacy concerns, particularly in relation to individual data, warrant

continuous attention and the development of robust privacy-preserving
measures

Real-world deployment and practicality considerations shed light on the

challenges and opportunities in integrating machine learning models into
law enforcement workflows. Scalability, user-friendliness, and
adaptability were identified as key factors influencing the successful
implementation of these tools.

The continuous improvement of models, their adaptability to evolving

crime patterns, and their positive impact on community trust and safety
are paramount for long-term success.

As we navigate the evolving landscape of crime prediction technology, it

is crucial to maintain a holistic perspective.

Balancing technical advancements with ethical considerations,

community engagement, and the realities of law enforcement operations
ensures that these tools contribute positively to public safety.

Future research and development should focus on refining models,

addressing ethical concerns, and fostering collaborations between data
scientists, law enforcement, and community stakeholders to create more
effective and responsible crime prediction systemsOnly through such
multidimensional efforts can we harness the potential of machine
learning for the betterment of society while safeguarding individual rights
and community well-being.
 Result:.

Features:

Enhanced Predictive Features:

Explore additional data sources and types of features that could further
improve the accuracy of crime prediction models. This might include
incorporating data from social media, urban infrastructure, or emerging
technologies such as Internet of Things (IoT) devices to capture more
nuanced patterns.
Temporal and Spatial Dynamics:
Investigate methods to better model temporal and spatial dynamics of
crime. Develop models that can adapt to seasonal variations, changes in
urban landscapes, and dynamic patterns of criminal activity over time.
Explainability and Interpretability:

Address the challenge of making crime prediction models more

interpretable and explainable. Future research could focus on
developing techniques to provide understandable insights to law
enforcement agencies and the public, increasing transparency and trust.
Human-Centric Design:

Incorporate user feedback and human-centric design principles to

ensure that crime prediction tools are practical and user-friendly for law
enforcement personnel. Consider the impact of these tools on decision-
making processes and develop interfaces that facilitate collaboration
between machine intelligence and human intuition.
Privacy-Preserving Techniques:

Explore advanced privacy-preserving techniques to address concerns

related to individual privacy. Research could focus on methods such as
federated learning, differential privacy, and secure multi-party
computation to allow model training without direct access to sensitive
individual data.
Cross-Domain Collaboration:

Encourage collaboration between researchers, law enforcement

agencies, policymakers, and communities to gather insights and
perspectives from diverse stakeholders. This collaboration can lead to
more comprehensive and effective crime prediction models that align
with societal needs.

Fairness and Bias Mitigation:

Develop and implement techniques for mitigating biases in crime

prediction models. Future works could explore algorithms that explicitly
address issues related to fairness, ensuring that predictions do not
disproportionately impact certain demographic groups.
Long-Term Impact Assessment:

Conduct studies to assess the long-term impact of deploying crime

prediction models on communities. Evaluate the effectiveness of these
models in reducing crime rates and enhancing public safety while
considering potential unintended consequences and ethical implications.
Dynamic Model Adaptation:

Explore methods for dynamic model adaptation that can quickly adjust to
changing crime patterns or unforeseen events. This could involve the
integration of real-time data and adaptive learning strategies.
Interdisciplinary Research:

Foster interdisciplinary research that combines expertise in criminology,

sociology, computer science, and data ethics. Collaboration between
experts in different fields can lead to a more holistic understanding of the
factors influencing crime and the development of more effective
prediction models.

Biblio graphy

Breiman, L. (2001). Random forests. Machine learning, 45(1), 5-32.

Powers, D. M. (2011). Evaluation: from precision, recall and F‐
measure to ROC, informedness, markedness and correlation. Journal of
Machine Learning Technologies, 2(1), 37-63.
Pedregosa, F., et al. (2011). Scikit-learn: Machine learning in Python.
Journal of Machine Learning Research, 12, 2825-2830.
Barocas, S., & Selbst, A. D. (2016). Big Data’s Disparate Impact.
California Law Review, 104(3), 671–732.
Diakopoulos, N. (2016). Accountability in Algorithmic Decision Making:
A Lens of Media Studies and Design. Data Society Research Institute.
 Hardt, M., Price, E., & Srebro, N. (2016). Equality of Opportunity in
Supervised Learning. In Advances in Neural Information Processing
Systems (NeurIPS).
Ratner, A., et al. (2017). Data programming: Creating large training sets,
quickly. Advances in Neural Information Processing Systems, 30.
Vapnik, V. (1995). The Nature of Statistical Learning Theory. Springer.
Thakkar, H., & Vondrick, C. (2020). Convolutional Neural Networks on
Non-uniformly Sampled Data. arXiv preprint arXiv:2006.07454.