A strategic framework is essential for categorizing emerging and highly domain-specific datasets, such as those in criminology and news. Classification is crucial for enhancing the usefulness and relevance of these datasets. This research introduces a framework that systematically classifies and structures complex domain-specific data, improving its utility for expert systems. By integrating the LLaMA (Large Language Model Architecture) with a lightweight machine learning classifier, XGBoost, the framework enhances the categorization process and enriches datasets with valuable information. The approach also includes the amalgamation of standard news APIs and uses the Cricket Algorithm for optimization with a diversity index to derive optimal solutions, focusing on improving precision, accuracy and F1 score in crime classification.
Classification and knowledge enrichment are critical to the advancement of theoretical research and practical applications in the specialized field of criminology. The deep relationships between many entities associated with crime are typically difficult to capture by traditional methods of data classification, owing to the dynamic and multifaceted character of criminal activity. The lack of comprehensive and precise crime data makes this problem worse and makes it more difficult to create efficient classification schemes. Integrating the dynamic, real-time data offered by news APIs is essential to meeting this problem. These APIs provide an ongoing flow of up-to-date information that, when paired with past crime data, allows for a more thorough and sophisticated assessment of crime trends. Researchers and practitioners can obtain important insights into trends, correlations, and developing patterns all crucial for efficient crime analysis and prevention by examining both historical and contemporary data. By introducing an innovative categorization system that integrates the newest generative AI technologies and machine learning techniques, the suggested framework seeks to close the gap in existing methodology.
The application of LLaMA (Large Language Model Architecture), which offers a solid foundation for semantic comprehension and classification, is essential to this strategy. The ability of LLaMA to synthesize specific information from several data sources improves the framework's capacity to analyze and classify intricate crime data. Moreover, Web 3.0 technologies are employed by the framework to enhance the classification procedure. Through a Sparkle-based engine to query news APIs, the system combines historical and current data, allowing partial knowledge graphs and ontological representations to be created. A more intuitive comprehension of the links and trends within the crime data is made possible by these knowledge graphs, which provide a visual and structured representation of the data. By utilizing the Cricket Algorithm for metaheuristic optimization and Shannon Entropy for feature selection, the framework guarantees the classification process's precision and effectiveness. This research enhances criminology and offers a useful tool for comprehending and preventing criminal activity by addressing the important demand for efficient classification and knowledge enrichment in the discipline.
Because criminal activity is dynamic and complex, it is becoming more important than ever to classify things in specialized fields like criminology. If properly categorized, patterns in both historical and modern crimes might provide insightful information for future study and useful applications. Nevertheless, current models are unable to represent the complex interactions between things because they lack the semantic depth and enrichment that come from contemporary APIs. By creating a solid, semantic-driven framework that makes use of Web 3.0 technologies and cuttingedge classification concepts, this research aims to close this gap. The proposed framework aims to advance the field and meet a critical need by integrating a sophisticated system of semantic similarity mechanisms to provide a more accurate and comprehensive approach to categorizing criminal activities based on historical and real-time data.
The proposed framework introduces several key contributions: it leverages LLaMA, a state-of-theart AI-generated language model, to enhance semantic understanding and classification of datasets. This framework also incorporates a novel dataset perspective using XGBoost, a lightweight machine learning algorithm, to effectively synthesize and classify features. By integrating advanced semantic AI techniques with dynamic background knowledge from APIs, and applying Simpson's Diversity Index and Shannon Entropy, the framework ensures diverse and robust classification outcomes. Additionally, the use of the Cricket Algorithm for metaheuristic optimization transforms feasible solutions into optimal ones. Finally, the framework not only classifies datasets with precision but also synthesizes knowledge representations in the form of trees or partially linked graphs, adding a distinctive semantic perspective that enhances the understanding of classified data.
The paper's organization is as follows: Section 2. comprehensively lists out a set of works related to this paper. Section 3 provides an overview of the proposed system architecture. Section 4 enumerates the results yielded by the proposed model. Finally, Section 5 concludes the paper.
A Bi-LSTM neural network was proposed by G. Deepak et al. [
Chhatwal et al. [
SS Kshatri et al. [
Shoeibi, et al. [
Jeyaboopathiraja et al. [
For getting current and pertinent news items, the Newsomatic API offers a comprehensive platform. With an emphasis on education, it provides a carefully chosen collection of news stories written in an approachable and captivating style for younger readers. Through the API, programmers can incorporate this content into apps to provide users with up-to-date news and instructional resources. The Newsomatic API has features like multimedia content, photos, and in-depth news stories that are all designed to promote learning and improve comprehension of current events. Applications can promote media literacy and awareness while providing richer news experiences by utilizing this API.
The British Broadcasting Corporation offers an abundance of news and multimedia content that may be accessed through the BBC API. Among the many services it provides are article retrieval, realtime news updates, and multimedia content including audio and video snippets. With the help of the API, developers can incorporate the BBC's comprehensive news coverage into their apps and provide customers reputable, high-quality news material. Access to breaking news, in-depth articles, and specialist information from several BBC news categories are among the features. Applications may provide their customers with dependable, varied news coverage by utilizing the BBC API, guaranteeing that users are informed about current events and stories throughout the world.
One of the top news organizations in the world, The Guardian, provides comprehensive access to its news and media content through the Guardian Open Platform API. With this API, developers can include news stories, video files, and interactive elements among other types of content from The Guardian's portfolio into their apps. Breaking news, in-depth articles, opinion pieces, and other editorial content can be retrieved via the API's endpoints. With features like subject classification, article metadata, and search capabilities, the Guardian Open Platform API enables applications to enhance user engagement with rich, carefully curated material from a reliable news source while delivering high-quality, current news content. 3.4. LLaMA LLaMA, an innovative advancement in artificial intelligence and natural language processing, was developed to extend the capabilities of machine learning models. It is predicated on the ideas of transformer architectures, which serve as the basis for models like BERT and GPT-4. However, the unique architectural enhancements and optimizations that make LLaMA unique enable it to handle massive data volumes and perform a wide range of language-related activities with astounding precision and effectiveness. Because it can be used for a wide range of tasks, from basic ones like text generation and translation to more complex ones like summarization and question answering, LLaMA is a versatile tool in the AI (Artificial Intelligence) arsenal. One of the most important aspects of LLaMA is its scalability. The model can handle bigger datasets and more parameters without requiring a proportional increase in processing capacity because of its efficient scaling. Because of the progress made in data processing and model training, LLaMA can function at a prominent level even when task complexity increases since it is scalable.
Additionally, the architecture of LLaMA incorporates strategies to mitigate common issues associated with big language models, such as the difficulty of preserving coherence in lengthier text outputs and the tendency to provide convincing but inaccurate information (hallucinations). Because of these developments, LLaMA is now a key player in the ongoing development of AI-driven language models. There are many different industries and fields in which LLaMA is used. When it comes to content creation, LLaMA can assist marketers and writers by creating unique content that is contextually relevant. This can boost productivity and save time. It can make it possible for customer support chatbots to understand and respond to user requests with accuracy, improving the customer experience. Furthermore, LLaMA can assist with data analysis, literature reviews, and the generation of original ideas in academic and research settings by analyzing and synthesizing enormous amounts of data. As LLaMA advances, it will further integrate into and transform a multitude of aspects of everyday and professional life, showcasing the immense potential of stateof-the-art AI technology.
XGBoost is a technology developed by Tianqi Chen and colleagues based on the gradient boosting approach, which iteratively boosts prediction accuracy by combining the strengths of numerous weak learners, usually decision trees. Thanks to its novel approach to model training and optimization, XGBoost can handle large-scale data sets and difficult prediction tasks with remarkable speed and performance. Owing to its capacity to predict customer behavior and identify fraudulent transactions, XGBoost has gained popularity across a variety of industries. XGBoost stands out from other boosting algorithms in part due to its innovative system architecture, which blends algorithmic optimizations with hardware-aware enhancements. The model efficiently handles sparse features and missing data by employing a novel sparsity-aware technique that is frequently observed in realworld data sets.
Furthermore, XGBoost uses second-order gradient information (Hessian) to perform the gradient descent process more precisely, leading to faster convergence and more accurate models. XGBoost is incredibly useful in applications where time is of importance because it can train models significantly faster than traditional gradient boosting approaches because of these technological advancements and parallel processing capabilities. The extensive use of XGBoost in academia and industry is evidence of its massive impact on the machine learning sector. It has consistently outperformed in numerous machine learning competitions, such as Kaggle, where it has proven essential to winning numerous challenges. Regression, classification, and ranking issues are just a few of the many applications that XGBoost excels at solving thanks to its scalability and robustness. Its open-source architecture and comprehensive documentation have also facilitated its integration with other machine learning pipelines and frameworks. XGBoost will continue to be excellent in modern machine learning approaches and become even more relevant and valuable as the demand for accurate and efficient prediction models grows.
The Shannon entropy, first proposed by Claude Shannon in 1948, quantifies the average quantity of information generated by a stochastic source by gauging the degree of uncertainty or unpredictability in a group of occurrences. It determines the average bit count needed to encode a message from a source optimally; higher entropy levels denote more complicated or ambiguous information.
( ) = − ∑ log2 , (1)
CoSimRank is a version of the SimRank method which was created to overcome some of its drawbacks and increase the effectiveness of similarity calculations in huge graphs. As the graph size grows, SimRank may become computationally costly; in contrast, CoSimRank provides a more scalable alternative. CoSimRank does this by using an iterative calculation technique that is more effective and decreases the temporal complexity involved in determining node similarities. CoSimRank offers a useful method for calculating similarities in large-scale networks by streamlining the recursive similarity calculation process. This makes it appropriate for use in recommendation systems, social network analysis, and web page ranking, among other uses. 3.8.
The Simpson's Diversity Index (SDI), which considers the total number of species present as well as their relative abundances, is a crucial indicator for evaluating biodiversity within a habitat. Ecologists and researchers can compare several habitats or monitor changes over time within a single habitat with the use of this index, which offers a quantifiable assessment of how diverse an ecosystem is. The Simpson's Diversity Index has a value between 0 and 1, where a number near 1 denotes minimal diversity, frequently because of a few dominant species, and a value of 0 suggests infinite diversity (with no single species dominating the community). Understanding the health of ecosystems, the state of conservation, and how environmental changes affect biodiversity all depend on this metric.
D = 1 − ∑ ni(ni−1),
N(N−1)
The Cricket Algorithm is an optimization method inspired by nature that solves challenging optimization issues by mimicking the behavior of crickets, especially their mating sounds. The algorithm was created by S. S. N. Singh and associates, and it was inspired by the way crickets use sound to find partners and create a mating network. This strategy is based on the ideas of swarm intelligence, in which unique actions support the process of overall optimization. Every cricket in the Cricket Algorithm stands for a fix for an optimization issue. The "calls" that the crickets make are symbolic of the caliber of their solutions. A cricket's attractiveness to others is determined by the frequency and volume of its calls. A cricket's fitness level affects the volume and frequency of its calls, increased fitness results in louder calls. This procedure is like how crickets utilize their sounds to entice mates and create a network of fixes. The two primary functions of the algorithm are exploration and exploitation. Crickets wander around the solution space in search of new locations during exploration and concentrate on fine-tuning solutions in areas of potential interest during exploitation. The secret to the Cricket Algorithm's success is its ability to achieve equilibrium between exploration and exploitation, which allows it to bypass local optima and converge towards superior solutions. It has been used in many different domains, including data mining, scheduling, and engineering design. Through the utilization of adaptive mechanisms and the application of cricket behavior, the Cricket Algorithm offers a reliable and adaptable solution for intricate optimization issues. Because of its ability to replicate natural processes and adjust to various issue environments, it is a useful supplement to the family of swarm intelligence algorithms. (2)
A single, sizable dataset created by combining many criminology and criminal news datasets was
used for the experiments. Among other datasets, the combined dataset contains the DICE (2022)
dataset, titled DICE: A Dataset of Italian Crime Event News [
The framework also employs a semantic approach using Shannon entropy and the Simpson's Diversity Index, calculated via the cricket algorithm a metaheuristic optimization technique. This ensures that the solutions generated are not only feasible but also optimal, contributing to the framework's robustness. The SSCKC is thus a comprehensive and strategically designed model that leverages lightweight machine learning, strong LLMs, and verified community news repositories to deliver a highly accurate and semantically enriched classification system. The CCPS (Conventional Crime Prediction System) model does not perform as expected primarily because it relies on individual machine learning models such as K-Means Nearest Neighbour, Decision Trees, Support Vector Machines, K-Nearest Neighbour, Extra Tree Classifier, and Artificial Neural Networks. These models are used in isolation rather than being hybridized in the framework. Without hybridization, these standalone models fail to deliver best-in-class results, as they lack the complementary strengths that a combined approach could offer. Moreover, the CCPS model does not incorporate Large Language Models (LLMs), which are crucial for generating contextual understanding and semantic enrichment of data. Additionally, the absence of community-verified facts, which serve as background knowledge in the SSCKC framework, further weakens the CCPS model's ability to provide accurate and reliable predictions. As a result, the CCPS model underperforms when compared to the proposed SSCKC framework, which benefits from a more sophisticated integration of machine learning, LLMs, and verified knowledge sources. The DLCCP (Deep Learning Crime Classification and Prediction) framework does not perform as expected because it relies solely on deep learning architecture for crime classification and prediction. While deep learning classifiers can be powerful, the model lacks semantics-oriented reasoning and learning, particularly in terms of quantitative reasoning and semantic similarity computation. This absence of semantic analysis diminishes the model's ability to understand and process crime data in a nuanced and context-aware manner.
Furthermore, the DLCCP framework does not incorporate authentic auxiliary knowledge from the Web 3.0 paradigm, limiting the depth and breadth of knowledge available to the model. This minimalistic knowledge base hinders the model's overall performance. Additionally, the framework does not utilize optimization techniques or nature-based metaheuristic models, which are essential for refining solutions and achieving optimal performance. As a result, the DLCCP framework underperforms when compared to the proposed SSCKC framework, which benefits from a more comprehensive approach that includes semantic reasoning, Web 3.0 knowledge, and advanced optimization methods. The CACP (Crime Analysis and Classification Platform) framework underperforms because it relies solely on two classifiers, Random Forest, and Support Vector Machine (SVM), without any advanced hybridization or integration with other models. The dataset used is basic, which limits the model's ability to handle more complex crime data.
Although the original article reports high accuracy, this drops significantly when applied to the proposed dataset due to the framework's lack of hybridization, absence of semantics-oriented augmented knowledge, and failure to include Large Language Models (LLMs) for caption generation or agents for query formalization. Most critically, the CACP framework does not leverage community-contributed, verified knowledge, which is essential in the proposed framework. These limitations result in CACP's inability to perform at the level of the more sophisticated SSCKC framework. For four models CCPS, DLCCP, CACP, and SSCKC the precision % is shown against the total number of recommendations in the graph. SSCKC has the highest precision at 95.12% of 10 recommendations, followed by CACP at 94.78%, DLCCP at 92.36%, and CCPS at 91.45%. SSCKC continues to be the best performer at 91.63% when recommendations reach 50, while CACP falls to 90.47%, DLCCP to 88.47%, and CCPS to 87.49%. This proves that SSCKC maintains the maximum precision at all recommended levels consistently.
The Proposed SSCKC, using a range of criteria such as false discovery rate (FDR), average precision, recall, accuracy, and F-measure. With an average precision of 97.12%, the Proposed SSCKC model outperforms the other models, which include CCPS at 89.90%, DLCCP at 90.75%, and CACP at 92.61%. SSCKC leads in average recall once more with 98.76%, followed by CCPS with 91.55%, DLCCP with 92.83%, and CACP with 93.76%. SSCKC has a higher average accuracy of 97.84% than CCPS (90.73%), DLCCP (91.79%), and CACP (93.19%). SSCKC scores a remarkable 98.03% on the F-measure, compared to scores of 90.71%, 91.80%, and 93.21% for CCPS, DLCCP, and CACP, respectively. Compared to CCPS (0.11), DLCCP (0.10), and CACP (0.08), SSCKC has the lowest false discovery rate (FDR), at 0.03. The Proposed SSCKC model is the most successful model in this comparison because it performs better overall in lowering false discovery rates and excels in precision, recall, accuracy, and F-measure.
This paper suggests a strategic framework for knowledge classification and generation specifically tailored for news and criminology. The framework uniquely combines the generative AI capabilities of LLaMA with the computational efficiency of the XGBoost machine learning algorithm, making it both cost-effective and practical. It integrates dynamic knowledge from current crime data and trends through news APIs, alongside historical data, to enhance its predictive capabilities. Utilizing a Sparkle-based engine for querying news APIs, Shannon Entropy for feature selection, and the Cricket Algorithm with Simpson's Diversity Index, the framework stands out for its ability to deliver high-quality, accurate solutions. This approach not only ensures precise crime classification but also enriches datasets with valuable knowledge, positioning the framework as a leading model in semantic crime and news classification with minimal false discovery rates.