2. Literature Review

In this section, the focus is on defining the research topic and establishing its significance within the broader academic discourse. By delving into past research endeavours, the literature review lays the groundwork for the subsequent stages of the study, providing context, theoretical frameworks, and empirical evidence to inform the research methodology and theoretical underpinnings. Furthermore, it offers insights into the evolution of ideas, methodologies, and perspectives over time, guiding in refining their research questions and hypotheses.

Expanding upon the literature review, a comparison between traditional methods and machine learning approaches in crime prediction reveals significant distinctions.

2.1 Traditional Methods vs Machine Learning

Traditional criminal activity forecast techniques mostly counts on analytical analysis and professional judgment which can have particular constraints compared to machine learning methods.

2.1.1 Limitations of Traditional Methods

Traditional criminal offense evaluation methods experience data sparsity, lack of interpretability, and their trust on human judgment which can be really prejudiced (Na et. al., 2021). Analytical designs may have a hard time to catch complicated patterns in crime information and might not adjust that well to advancing criminal offense trends. As stated, traditional techniques often deal with obstacles with information sparsity which indicates that they might not have enough pertinent data indicate make precise forecasts. This can cause unreliable outcomes (Bennell et. al., 2009). Parametric statistical designs which are used extremely widely lack interpretability which makes it difficult for researchers and practitioners to recognize the primary mechanisms driving their predictions (Noraga, 2021). For instance, consider a study where researchers used logistic regression to predict the likelihood of heart disease based on various risk factors like age, gender, cholesterol levels, and blood pressure. While the model might accurately predict the probability of heart disease, it may be challenging for researchers and practitioners to understand how each predictor variable contributes to the overall prediction (Ismail and Anil, 2014). This lack of openness can limit their worth when it pertains to making plan decisions. There is also an absence of interpretability when it involves typical techniques. Standard techniques commonly rely upon subjective human judgment for version growth and interpretation. This presents bias and can lead to inconsistent results. This can occur particularly in the situations when different specialists are associated with the analysis procedure (Norga, 2021). Analytical models might battle to obtain intricate patterns which could be present in the crime data, such as nonlinear partnerships or communications among [unreadable] variables.

2.1.2 Advantages of Machine Learning

Machine learning algorithms can efficiently process large amount of data from multiple sources which includes demographic information, socioeconomic factors, and historic crime data. This capability allows for a more complete analysis of crime patterns and trends. Machine learning algorithms are good at identifying complex patterns and relationships within the data that may not be clear by doing manual analysis (Jenga et. al., 2023). By getting meaningful patterns, these algorithms can find insights that help in understanding the underlying factors which lead to criminal behaviour. Through the process of training on historical data, machine learning models learn to recognize patterns connected with different types of criminal activities. This information helps them make right predictions about future crime occurrences, helping law enforcement agencies allocate resources more effectively. Machine learning algorithms can analyse complex interactions among various factors influencing crime occurrence, such as socioeconomic conditions, environmental factors, and law enforcement activities (Jenga et. al., 2023). By considering these interactions, machine learning models can provide more accurate predictions than traditional methods.

2.1.3 Comparative Analysis

There have actually been studies done which contrasts the use of standard techniques. It mostly shows the prevalence of machine learning approaches in terms of prediction precision and versatility. Machine learning formulas, such as semantic networks and arbitrary forests can evaluate complicated relationships in criminal activity information which leads to extra precise predictions compared to typical designs (Mandalapu et. al., 2023). The versatility of artificial intelligence models is why it is understood to be much better than conventional methods. Artificial intelligence models are capable of adjusting to transforming criminal activity patterns and progressing datasets, making them more versatile in real-world scenarios (Shah et. al., 2021). In contrast, typical techniques usually require manual changes or recalibration, which can be lengthy and might not appropriately capture the altering dynamics of the data. Machine learning designs also have the capacity to handle huge amounts of datasets that make them better (Taye, 2023). An additional advantage of machine learning is its capacity to manage diverse information resources. Artificial intelligence versions can in fact integrate organized and unstructured information from different resources, such as message and pictures to enhance forecast accuracy (Durap, 2023). This capability to use a selection of data enables artificial intelligence algorithms to find covert insights and patterns that might not be clear when using typical approaches. It is not like AI is better than traditional methods, but AI along with machine learning algorithms do have an upper hand.

2.2 Types of Machine Learning Algorithms

There are several machine learning algorithms which are and can be used in crime prediction such as decision trees, random forests, support vector machines (SVM), and neural networks. Each algorithm has its strengths and weaknesses which impacts prediction accuracy and computational efficiency (Jenga et. al., 2023).

2.2.1 Decision trees

Choice trees are hierarchical structures that split information right into branches based on functions. They are very easy to translate and imagine, making them beneficial for comprehending the decision-making process (Jenga et. al., 2023). Decision trees work extremely well when it pertains to creating prediction designs for numerous aspects of crime, such as crime reporting, transgressor profiling, and criminal activity event (Gutierrez and Leroy, 2007). Choice trees help identify the most vital features contributing to criminal activity prediction. Analysing the framework of the choice tree can help obtain a concept about the aspects which have one of the most substantial effect on criminal behaviour (Shi et al., 2018). This tells that choice trees can resolve the major concern of recognizing and comprehending features that would be connected with any type of sort of neighbourhood criminal offence. One of the essential advantages of choice trees in criminal activity prediction is their interpretability. Interpretability is just the possibility of locating meaning behind something. Law enforcement agencies and policymakers can quickly understand the decision policies created by the design, aiding in decision-making procedures (Shah et. al., 2021). Decision trees have actually been used as a relied on algorithm and has been previously made use of to forecast offenders' indigenous areas based on example information. This application has helped law enforcement agencies narrow down their search locations, potentially leading to quicker apprehension of crooks (Shi et. al., 2018). It is pretty evident that decision trees play a substantial function in predicting criminal offence as a result of their ability to take care of both categorical and numerical information properly. Choice trees and commonly contrasted to random woodland while reviewing concerning the much better algorithm to predict neighbourhood criminal offence.

Using decision trees for crime prediction involves constructing a hierarchical structure where each node represents a test on an attribute, each branch represents the test outcome, and each leaf node signifies a class label (crime occurrence or non-occurrence). This approach entails data pre-processing, feature selection, training on historical crime data, and application to new datasets. Decision trees offer interpretability, aiding stakeholders in understanding crime prediction factors easily (Blum and Langley, 1997). They are user-friendly, facilitating quick analyses and data exploration (Oleszak, 2023). Moreover, decision trees handle non-linear relationships effectively, capturing intricate patterns in crime data akin to random forests.

However, decision trees have drawbacks. They are prone to overfitting, especially with deep trees, necessitating pruning and hyperparameter tuning (Brownlee, 2020). Additionally, they can be unstable, leading to significant variations in predictions due to minor data changes; this can be mitigated with ensemble methods (Leroy and Gutierrez, 2007). Furthermore, decision trees may struggle to capture complex relationships, impacting predictive accuracy (Leroy and Gutierrez, 2007).

2.2.2 Random Forest

Random forests are nothing but groups of choice trees that improve prediction precision by combining multiple trees' outputs. They deal with noisy data well and are much less prone to overfitting (Shah et. al., 2021). Here, each tree is trained on a random subset of the information and features, which subsequently ensures variety in the specific trees (Adnan and Islam, 2017). By doing so and incorporating the outputs of several choice trees, arbitrary woodlands tend to

achieve greater prediction accuracy compared to private trees. This consequently helps to reduce the distinction and bias existing in any type of single tree, leading to even more proper predictions (Boehmke and Greenwell, 2020). One more plus factor concerning random woodland is that they deal extremely well with noisy information in addition to outliers. This quality makes arbitrary forest suitable for uncertain datasets such as the regional criminal activity dataset (Boehmke and Greenwell, 2020). As it is apparent that the CSEW regional crime dataset can be substantial, so the arbitrary woodland algorithm is one of the most helpful one in this circumstance as it is suitable to manage large information. In general, random woodlands provide a powerful and flexible device discovering technique for numerous forecast jobs, consisting of classification and regression, because of their capacity to make use of and harness the toughness of numerous choice trees while lessening their weaknesses.

Using a random forest model for predicting crime involves employing machine learning techniques to analyse historical crime data and make predictions about future crime occurrences. This method typically involves pre-processing the data, selecting relevant features, training the model on historical crime data, and then using the trained model to predict crime in new datasets.

There are certain strengths and advantages of random forest which make it suitable to predict crime including:

High Accuracy: Random forests exhibit high accuracy by leveraging ensemble learning, aggregating predictions from multiple decision trees. This approach helps capture intricate patterns and dependencies in crime data, leading to more accurate predictions (Mandalapu et. al., 2023).

Handles Non-linear Relationships: Unlike linear regression, which assumes linear relationships between variables, random forests can effectively model non-linear relationships. This capability allows them to capture complex interactions between crime predictors, such as demographics, environmental factors, and past crime rates, enhancing predictive performance (Mumuni and Mumuni, 2022).

Feature Importance: Random forests provide insights into feature importance, indicating which variables contribute most significantly to crime prediction. Understanding feature importance aids law enforcement agencies and policymakers in prioritizing interventions and allocating resources effectively to prevent and address crime (Brick et. al., 2018).

Along with strengths, there comes limitations such as:

Overfitting: Random forest models are susceptible to overfitting, especially when the number of trees in the forest is too high. Additionally, improper tuning of hyperparameters can exacerbate this issue, leading to models that perform well on training data but generalize poorly to unseen data (Scholau and Zou, 2020).

Interpretability: Despite their high accuracy, random forest models can be challenging to interpret. Unlike simpler models like linear regression, which provide clear coefficients for each predictor, random forests lack transparency in the decision-making process. Understanding how individual features contribute to predictions can be complex, limiting the model's interpretability (Khan et. al., 2022).

Computationally Intensive: raining random forest models can be computationally intensive, particularly with large datasets. The process of growing multiple decision trees and aggregating their predictions requires significant computational resources, potentially leading to long training times and increased computational costs (Tollenaar and Heijden, 2019).

2.2.3 K Nearest Neighbours (KNN)

K-Nearest Neighbours (KNN) is a popular machine learning algorithm used in crime prediction due to its simplicity and effectiveness in classification tasks. KNN is a non-parametric algorithm that classifies a new data point based on the majority class of its k nearest neighbours in the feature space. It does not require explicit training and can adapt to complex decision boundaries (Jenga et. al., 2023). Another application of KNN is its spatial analysis. KNN's reliance on spatial proximity makes it well-suited for spatial crime analysis. By considering the geographical locations of crime incidents, KNN can identify spatial clusters and hotspots, aiding law enforcement agencies in allocating resources effectively (Shah et. al., 2021).

Notable Studies Using KNN for Crime Prediction:

1. Study by Saeed et al.: Investigated the use of supervised learning approaches, including KNN, for crime prediction, highlighting its prevalence in the literature (Saeed and Abdulmohsin, 2023)

2. Systematic Literature Review by Shah et al.: Provided insights into the application of data mining techniques, including KNN, in crime prediction research (Shah et. al., 2021)

3. Research by Mandalapu et al.: Explored the use of KNN and other machine learning algorithms in crime prediction, contributing to the advancement of predictive models (Mandalapu et al., 2023).

These studies have extensively explored the application of KNN and other machine learning algorithms in crime prediction, shedding light on its prevalence in the literature and contributing to advancements in predictive models. The utilization of K-Nearest Neighbours (KNN) in crime prediction is underscored by its simplicity, effectiveness in classification tasks, and adaptability to complex decision boundaries, as highlighted by Jenga et al. (2023). However, while KNN offers several advantages, its performance may be hindered by challenges such as data imbalance and the need for parameter tuning. Thus, future research should focus on addressing these limitations to maximize the utility of KNN in crime prediction applications.

and trends in crime patterns. However, the choice of the optimal value for k and the selection of relevant features are critical in ensuring the accuracy and reliability of the predictions (Jenga et al., 2023; Kaufmann et al., 2018). KNN offers several strengths and limitations.

Simplicity is one of KNN's primary strengths. It is straightforward to understand and implement, making it accessible for beginners in machine learning. Its intuitive concept of classifying data points based on their proximity to neighbouring points makes it easy to grasp and apply in various scenarios (Zhang, 2016). Additionally, KNN is non-parametric, meaning it does not assume any underlying data distribution. This flexibility allows KNN to handle various data types and adapt to different types of datasets without requiring prior assumptions about the data's distribution (Guo et al., 2003). Another advantage of KNN is its lack of a training period. Since KNN utilizes the entire dataset as its model, there's no separate training phase required. This absence of a training period reduces pre-processing time and simplifies the workflow, especially for datasets where real-time predictions are necessary (Zhang, 2016). Moreover, KNN demonstrates adaptability, performing well in scenarios with a large number of classes and effectively handling irregular or complex decision boundaries. Its ability to adapt to different data distributions and learn from local patterns makes it a versatile algorithm suitable for various classification tasks.

However, KNN has its limitations. One drawback is its computational expense. KNN involves calculating distances between query points and all other data points, which can be computationally intensive, especially for large datasets. This computational complexity increases with the size of the dataset, leading to longer processing times. Additionally, KNN is memory-intensive, as it stores all training data in memory. This can lead to memory constraints, particularly with large datasets, where the memory required to store all data points increases proportionally. Furthermore, KNN is sensitive to noise, outliers, and noisy data. Outliers can significantly impact the algorithm's performance by influencing the proximity-based classification, leading to less accurate predictions. Therefore, data preprocessing steps such as outlier detection and removal are crucial when using KNN (Zhang, 2016). Lastly, KNN is not suitable for high-dimensional data. Its performance diminishes as the number of dimensions or features increases due to the curse of dimensionality. With high-dimensional data, the distance between data points becomes less meaningful, making it challenging for KNN to accurately identify nearest neighbours and resulting in reduced performance.

2.2.4 Logistic Regression

Logistic regression is a statistical method commonly employed in predictive modeling to examine the relationship between a dependent variable, such as the likelihood of a crime occurring, and one or more independent variables, such as demographic factors, socioeconomic indicators, or historical crime data. Unlike linear regression, logistic regression is well-suited for binary classification tasks, making it ideal for predicting binary outcomes, such as the presence or absence of a crime event. It estimates the probability of an event occurrence by fitting data to a logistic function, transforming the linear regression output into a range between 0 and 1, representing probabilities. Instances with probabilities above a certain threshold are classified into one class, while those below are classified into the other. Logistic regression is favored in crime prediction tasks due to its simplicity, interpretability, and efficiency (Hosmer Jr et al., 2013) (Wickham et al., 2019).

The strengths of logistic regression include its interpretability, as the coefficients of logistic regression provide insight into the direction and magnitude of predictors' influence on the outcome variable. This enables straightforward interpretation of the model's impact, making it accessible even to non-technical stakeholders. Moreover, logistic regression is computationally efficient, making it suitable for large datasets and real-time prediction applications. Unlike more complex models, logistic regression does not require extensive computational resources, allowing for faster model training and prediction. Additionally, logistic regression provides probabilistic outputs in the form of predicted probabilities, facilitating the estimation of confidence levels in predictions. This probabilistic nature is particularly useful in decision-making scenarios where uncertainty assessment is crucial. Furthermore, logistic regression is robust to small amounts of noise and irrelevant features in the dataset, effectively filtering out irrelevant variables and focusing on those with significant predictive power, ensuring reliable model performance even in imperfect data environments (Sperandei, 2014).

However, logistic regression also has limitations. Firstly, it assumes a linear decision boundary, which can restrict its ability to accurately model complex relationships between predictors and the outcome. In scenarios where relationships are non-linear, logistic regression may struggle to capture underlying patterns effectively. Additionally, logistic regression relies on the assumption of linearity between predictors and the log odds of the outcome, which may not hold true in all cases, especially when relationships are non-linear or more complex. Moreover, the performance of logistic regression heavily depends on feature selection and engineering, requiring domain knowledge and expertise to choose relevant predictors. The quality of selected features significantly impacts the model's predictive performance. Lastly, logistic regression may encounter challenges with imbalanced datasets, where one class significantly outweights the other. In such cases, the model may exhibit biased predictions, favoring the majority class and leading to suboptimal performance, especially when minority class samples are crucial (Sperandei, 2014).

In conclusion, logistic regression is a valuable tool in crime prediction tasks, offering simplicity, interpretability, and efficiency. However, it's essential to consider its limitations and suitability for specific datasets and scenarios.

2.3 Challenges and Limitations

Criminal activity forecast making use of machine learning provides significant potential for boosting law enforcement approaches. Nevertheless, in addition to advantages comes obstacles. There are a number of challenges when it involves criminal offense prediction making use of artificial intelligence. These difficulties and restrictions require to be addressed in order to ensure the efficiency of forecasts. Such difficulties consist of:

2.3.1 Data Quality Issues

The top quality of the input information is what issues when it's about the success of machine learning designs. In criminal activity prediction, datasets might suffer from mistakes, incompleteness, or predispositions, which can compromise the dependability of predictive models (Saeed and Abdulmohsin, 2023). Mistakes are a significant problem as crime information might include errors because of misreporting, misclassification, or variances in information collection methods (Pina-Sánchez et. al., 2022). These errors can bring about

wrong understandings and forecasts if not handled correctly. Additionally, crime datasets might not have particular pertinent information or experience missing data, which can influence the efficiency of predictive models (De Moor et. al., 2020). This can bring about the missing out on information presenting bias and decreasing the precision of forecasts. Resolving these information quality problems requires complete information cleansing, normalization along with recognition processes to make certain the accuracy and integrity of the data which is utilized for training and screening predictive formulas (Purves, 2023). Data cleansing includes recognizing and correcting errors, inconsistencies, and missing worths in the dataset. Normalization guarantees that data which has actually been gathered from different sources are also and compatible for evaluation. Recognition techniques, such as cross-validation action the performance and reliability of anticipating versions using independent datasets. By doing all these, we can enhance the accuracy, integrity, and legitimacy of machine learning models for criminal activity prediction.

2.3.2 Interpretability of Models

Some complex deep learning algorithms are often seen as “black boxes” which makes it very difficult for policymakers to understand how predictions are made (Saeed and Abdulmohsin, 2023). Lack of model interpretability affects trust and transparency in predictive policing systems which in turn raises concerns about accountability and potential biases (Linardatos et. al., 2021). This problem can be solved by using interpretable machine learning models. This can help to understand the decision- making process much better. By explaining how predictions are generated, interpretability techniques bridge the gap between the predictive power of machine learning models and peoples' understanding of those predictions. Efforts to enhance model interpretability involve developing explainable AI techniques that provide insights into the decision-making process of machine learning algorithms. These efforts can definitely contribute to the development of more transparent, fair, and accountable systems for crime prediction and prevention.

To enhance machine learning model interpretability, several methods can be employed:

1. Application-grounded Evaluation: Assess interpretability based on how well it serves the application's needs and objectives (Alangari et. al., 2023). This method ensures that interpretability aligns with practical needs, enhancing its relevance. However, it may overlook broader interpretability aspects crucial for generalizability.

2. Human-grounded Evaluation: Involve human judgement to determine the effectiveness (Linardatos et. al., 2020). Involving human judgement acknowledges the importance of user perception. Yet, it introduces subjectivity and potential bias in the evaluation process.

3. Functionally-grounded Evaluation: Evaluate interpretability by examining how well it enables users to comprehend the model's internal functioning and decision-making processes (Molnar, 2023). Focusing on the comprehension of internal model processes provides valuable insights. However, it may overlook broader interpretability aspects crucial for generalizability.

4. Rule-based Models: Employ simple rule-based models that provide explicit rules for decision-making, enhancing transparency and understanding. Rule-based models offer

transparency and simplicity, aiding interpretability. Yet, they may lack the flexibility to capture complex relationships in data.

5. Local Interpretable Model-agnostic Explanations (LIME): Utilize techniques like LIME to explain individual predictions of complex models, offering insights into model behaviour at the instance level (Salih, 2022). LIME provides detailed insights into individual predictions, enhancing transparency. However, it may not capture the overall model behaviour, limiting its applicability.

By employing these methods, machine learning practitioners can improve the interpretability of models, fostering trust, collaboration, and informed decision-making.

2.3.3 Overcoming Bias

Bias in machine learning formulas causes several challenges in criminal offense prediction, specifically as a result of predispositions which are built-in in historic criminal offense data (Giffen at. al., 2022). It can cause biased end results and highlight the existing inequalities in law enforcement methods. To solve this difficulty, it is important to collect data from numerous resources. Addressing this problem likewise needs cautious feature choice which is essential to criminal activity forecast (Jenga et. al., 2023). Chatting with professionals can assist with this process. Another point which can be done is to develop formulas that are durable to prejudices and sensitive to fairness considerations. Methods such as making use of fairness controls into the optimization procedure to decrease the bias have actually been encouraging (Ferrara et. al., 2024). By integrating all these methods, it can be fairly simple to get over any kind of prejudice in any type of equipment discovering version.

In summary, the application of machine learning in criminal activity prediction offers substantial potential for enhancing law enforcement strategies. However, several challenges must be addressed to ensure the effectiveness and ethics of predictive models. By addressing these challenges and implementing collaborative governance approaches tailored to

2.3.4 Collaborative Governance

When it pertains to anticipating regional crime and its facets, collective administration plays an essential role in making sure the efficiency and ethicality of the predictive policing efforts. Engaging with regional communities in the prediction of criminal activity develops count on and it also guarantees that policing efforts align with the area needs and values (Ruijer, 2021). There is a requirement to establish data collaboratives as it involves collaborations between police, government entities, and neighbourhood organizations. This aids in ensuring that the data is shared properly which consequently enhances the precision of criminal activity prediction versions. Including numerous people also guarantees that multiple varied viewpoints are thought about, and hence potential biases are dealt with (Polzin and Wilson, 2014). Something which can build public self-confidence regarding any type of predictive policing campaign is applying transparent procedures (Kaufmann et. al., 2018). Hence, by taking on collective governance approaches made according to the local scenarios, anticipating policing initiatives can be a lot more efficient, neutral, and responsible. This can assist keep much safer communities.

In summary, the application of machine learning in criminal activity prediction offers substantial potential for enhancing law enforcement strategies. However, several challenges must be addressed to ensure the effectiveness and ethics of predictive models. By addressing these challenges and implementing collaborative governance approaches tailored to

local contexts, predictive policing initiatives can be more effective, impartial, and accountable,

ultimately contributing to safer communities.

2.4 Case Studies

Several successful implementations of machine learning-based crime prediction systems have been previously observed in various jurisdictions, using crime datasets to understand local crime dynamics and enhance law enforcement strategies.

2.4.1 London Metropolitan Police

The London Metropolitan Police made use of machine learning algorithms to analyse vast amounts of historical crime data, demographic information, and other relevant variables. By employing advanced predictive analytics, they were able to identify patterns and trends indicative of areas with high crime rates (Reese, 2022). This approach enabled law enforcement to allocate resources strategically, focusing efforts on areas predicted to experience increased criminal activity. Additionally, machine learning models provided insights into the underlying factors contributing to crime which led to developing targeted crime prevention initiatives and programs. The London Metropolitan Police have experimented with machine learning for crime prediction mainly through facial recognition technology (Reese, 2022). However, it is a fact that such methods can be prone to inaccuracies. Trials were conducted around this method using deep learning algorithms, but its effectiveness has always been a concern.

2.4.2 Chicago Police Department

The Chicago Police Department implemented predictive policing models using machine learning algorithms to forecast crime hotspots, resulting in a decrease in violent crimes within targeted regions. One such model which was developed by this police department is called as the Strategic Subjects List (SSL) which uses machine learning algorithms to identify individuals at a higher risk of being involved in violent crimes, either as victims or criminals (Coates, 2017). This will definitely help to prevent any future crime and therefore enhances the overall public safety. More initiatives were undertaken in Chicago such as the Crime Prevention and Information Center (CPIC) which used data analytics and predictive modelling to predict criminal activity, enabling law enforcement to respond quickly and prevent violent incidents (Tamir et. al., 2021). This has successfully demonstrated that the use of machine learning algorithms in predictive policing does help the community. But the downside of this is that some studies have also shown that there is a potential for racial discrimination and biases when it comes to predictive policing algorithms (Hung and Yen, 2023) which of course can impact the fairness of the policy.

2.4.3 Los Angeles Police Department

The Los Angeles Police Department have also implemented a predictive analysis system which combined the crime data and the socioeconomic factors to identify areas at higher risk for specific types of crimes. They have used big data for their policing and since then it has been a topic of discussion and examination about how technologies like predictive analytics impact

surveillance practices and law enforcement strategies (Brayne, 2017). There have been several

studies conducted to understand the impact of such predictive analytics systems on surveillance

practices and law enforcement strategies. The effectiveness, ethical considerations, and

potential biases associated with predictive policing technologies were analysed (Heaven,

2020). It can be analysed that the Los Angeles Police Department’s use of predictive analysis

shows the intersection of technology, data, and law enforcement.

2.4.4 New York Police Department

The New York Police Department used machine learning tools to enhance crime analysis and

pattern recognition (Wood, 2019). The algorithms were implemented to analyse crime data.

They help analysts and officers in identifying relationships within crime data in turn helping in

the detection of crime patterns. Through the use of predictive analytics, the NYPD optimized

patrol routes and resource allocation strategies, leading to a reduction in overall crime rates

throughout the city. The NYPD's algorithm, Patternizr, is designed to track crimes citywide

and recognize emerging patterns (Charles, 2019). Such initiatives aim to improve law

enforcement's ability to allocate resources effectively and address criminal activity. However,

there have been certain concerns about AI bias and its usage (Holak, 2019). Despite this, using

machine learning for this does present promising opportunities for crime prediction and its

analysis. It's crucial for law enforcement agencies to ensure transparency, accountability, and

fairness in the development and deployment of machine learning algorithms to lessen biases

and maintain civil rights.

2.5 Perspectives on Crime Prediction

There are several perspectives when it comes to crime prediction and they refer to the several knowledge frameworks and methodologies which are used in those crime prediction practices. Such frameworks include mathematical, social and technological.

2.5.1 Mathematical Framework

Mathematical framework and methodology in crime prediction involves using quantitative analysis and statistical modelling techniques to uncover patterns and trends within crime data. This approach relies on mathematical algorithms and computational methods to process large volumes of data efficiently (Hälterlein, 2021). By applying mathematical principles, such as probability theory and regression analysis, analysts can identify correlations between various factors and criminal activities.

One key aspect of mathematical framework is its focus on predictive modelling. By gathering information from historical crime data, mathematical models can forecast future criminal activities in specific areas or among particular demographics. These models can inform law enforcement agencies and policymakers about potential hotspots for criminal behaviour, enabling proactive intervention strategies (Modise, 2023).

Moreover, it also provides a systematic framework for evaluating the effectiveness of crime prevention measures. By comparing predicted outcomes with actual occurrences, analysts can assess the accuracy of their models and refine them accordingly (Hälterlein, 2021). This process contributes to the continuous improvement of predictive capabilities in crime prevention efforts. There can be certain potential limitations and challenges associated with this approach. These may include the inherent biases present in historical crime data, the complexity of human behaviour, and ethical considerations surrounding the use of predictive analytics in law enforcement. Additionally, there may be concerns regarding the transparency and interpretability of mathematical models, which could impact their acceptance and effectiveness in practice.

Overall, while mathematical frameworks offer valuable tools for understanding criminal behaviour and guiding preventive measures, a better approach is necessary to address the complexities and limitations inherent in crime prediction methodologies.

2.5.2 Social Framework

The methodologies which concern the social part of crime prediction delves into the intricate relationship between various socio-economic, cultural, and environmental factors that contribute to criminal behaviour. It examines how societal structures, economic disparities, cultural norms, and environmental conditions influence individuals' tendency to engage in criminal activities (Hälterlein, 2021). By examining these dynamics, social framework aims to understand the actual mechanisms driving crime within communities. To inform crime prediction strategies effectively, social epistemology emphasizes the importance of comprehensively understanding the social fabric and relationships within communities. This involves studying community dynamics, social networks, group interactions, and institutional factors that shape individuals' behaviours and decisions regarding crime involvement.

There can be certain concerns regarding social framework where one concern includes the complexity of accurately capturing and quantifying the diverse array of social factors that contribute to criminal behaviour. Additionally, the dynamic nature of social dynamics and their interaction with individual choices pose challenges in developing predictive models that account for temporal and contextual variations. Also while the social framework focuses on the importance of understanding community dynamics and social relationships, there may be limitations in operationalizing this knowledge into actionable crime prediction strategies.

2.5.3 Technological Framework

Technological framework refers to the application of advanced technologies, particularly artificial intelligence (AI), machine learning (ML), and big data analytics, in understanding and predicting complex phenomena such as criminal behaviour (Alvarado, 2023). Here's an expansion of the concept:

• Artificial Intelligence (AI): AI encompasses various techniques aimed at creating systems capable of performing tasks that typically require human intelligence. In crime prediction, AI algorithms analyse large datasets to identify patterns and trends, aiding in the anticipation of criminal activities (Dakalbab et. al., 2022).

• Machine Learning (ML): ML algorithms enable systems to learn from data without being explicitly programmed. In crime prediction, ML models can analyse historical crime data to identify factors contributing to criminal behaviour and make predictions about future incidents (Mandalalapu et al., 2023).

• Big Data Analytics: Big data analytics involves the exploration and analysis of large and complex datasets to uncover hidden patterns, correlations, and other insights. In crime prediction, big data analytics can process diverse sources of information, including social media, sensor data, and law enforcement records, to generate actionable intelligence (Feng et. al., 2019).

• Enhanced Predictive Insights: By harnessing these technologies, technological epistemology aims to provide law enforcement agencies with more accurate and timely predictive insights into crime trends and hotspots. This facilitates proactive measures to prevent crime and allocate resources effectively (Hussein and Abdulameer, 2022).

It is now evident that technological framework also has several strengths and limitations where the strength is the integration of AI, ML, and big data analytics which offers promising opportunities to analyse extensive datasets and identify intricate patterns in criminal behaviour, enabling proactive crime prevention strategies. On the other hand, the limitation is that reliance only on technological solutions may overlook the social, economic, and cultural factors influencing crime. Human biases in data collection and algorithmic decision-making could continue disparities and inaccuracies in crime prediction models.

Overall, technological framework plays a crucial role in advancing crime prediction capabilities by using cutting-edge tools and techniques to analyse vast amounts of data and generate actionable intelligence.

In conclusion, this literature review has provided a comprehensive analysis of the existing research on prediction of local crime using machine learning. By synthesizing and evaluating the findings from various studies, it has illuminated key themes, trends, and gaps in the current knowledge landscape. Moving forward, this review sets the stage for future research endeavours aimed at addressing these gaps and advancing our understanding of crime prediction. The identified research gaps suggest avenues for further investigation, such as exploring the impact of various crime types in various areas, investigating the efficacy of

machine learning algorithms in predicting crime along with comparing the accuracies of

individual algorithms.