INDIAN CRIME DATA, 2020

The framework analyses data sources, employ data pre-processing techniques, apply machine learning algorithms, incorporate investigative support to enhance crime detection and prediction capabilities. It encompasses data collection, pre-processing, exploratory data analysis, feature selection and engineering, crime detection, crime prediction, investigation support, deployment and monitoring, as well as collaboration and knowledge sharing. By collecting relevant data from multiple sources crime reports, arrest records, footage it ensures a comprehensive dataset. Data pre-processing techniques are employed to clean, normalize, and transform the collected data. Exploratory data analysis provides insights into crime patterns, trends, correlations. Feature selection and engineering help identify the most relevant features for crime detection and prediction. Clustering algorithms are utilized to identify spatial patterns and crime clusters. It also incorporates investigation support by utilizing techniques to identify relationships, associations, and networks between individuals, locations, and events. Deployment and monitoring ensure its integration into operational systems for real-time or batch processing. The framework encourages collaboration and knowledge sharing between law enforcement agencies, researchers, and data scientists. The proposed Spatio-Temporal Contextual Crime Predictor framework in this research for investigating, detecting, and predicting crime using data mining is built upon a robust, multi-iteration pipeline. This pipeline meticulously extracts and preprocesses structured tabular crime data from complex, multi-page PDF documents, addressing the challenges of India's National Crime Records Bureau data, often non-machine-readable.

For crime detection and prediction, the STCCP framework primarily employs Random Forest Regressor for forecasting crime rates and identifying potential hotspots, building a model that predicts crime incidence.

A key novelty of STCCP is its contextual feature enrichment via Large Language Models (LLMs), which transform raw, disparate data into explainable, indexed knowledge units, providing richer narrative context and abstract indicators vital for comprehensive crime analysis. This fusion of LLM-derived context with geospatial and temporal modelling enables interpretable decisions, a crucial advantage over traditional "black-box" approaches.
Investigation support is provided through the application of the trained Random Forest model, which aids in identifying patterns and extracting interpretable decision rules from crime data. The framework effectively incorporates Random Forest as a core classification and prediction tool, leveraging data mining techniques to investigate, detect, and predict crimes. The STCCP consistently outperforms state-of-the-art models in predictive accuracy, enhancing crime-fighting capabilities, improving resource allocation, and contributing to safer communities.

该框架对多源数据开展分析，采用数据预处理技术，应用机器学习算法，并整合侦查辅助手段，以提升犯罪侦查与预测能力。其涵盖数据采集、预处理、探索性数据分析、特征选择与工程化、犯罪侦查、犯罪预测、侦查辅助、部署与监控，以及协作与知识共享等全流程环节。

通过从犯罪报告、逮捕记录、监控影像等多源渠道收集相关数据，可构建覆盖全面的数据集。数据预处理技术用于对采集到的数据进行清洗、归一化与转换操作。探索性数据分析可揭示犯罪模式、趋势与关联关系。特征选择与工程化环节有助于筛选出适用于犯罪侦查与预测的核心相关特征。研究采用聚类算法以识别空间分布模式与犯罪集群。

该框架还通过相关技术整合侦查辅助功能，用于识别个体、地点与事件之间的关联关系与网络结构。部署与监控环节可确保该框架集成至业务系统，实现实时或批量数据处理。该框架推动执法机构、研究人员与数据科学家之间的协作与知识共享。

本研究提出的时空上下文犯罪预测（Spatio-Temporal Contextual Crime Predictor, STCCP）框架，依托稳健的多迭代流程，用于借助数据挖掘技术开展犯罪侦查、检测与预测工作。该流程可从复杂的多页PDF文档中精准提取并预处理结构化表格犯罪数据，攻克了印度国家犯罪记录局（National Crime Records Bureau, NCRB）数据通常难以被机器读取的技术难题。

在犯罪侦查与预测任务中，STCCP框架主要采用随机森林回归器（Random Forest Regressor）预测犯罪率并识别潜在犯罪热点，构建可预测犯罪发生情况的模型。

STCCP的核心创新之一在于借助大语言模型（Large Language Models, LLMs）实现上下文特征富集：该技术可将原始异构数据转化为可解释的索引化知识单元，为全面的犯罪分析提供更丰富的叙事语境与抽象指标。将LLM衍生的上下文信息与时空建模相结合，可实现可解释的决策过程，这相较于传统的“黑箱”模型具有显著优势。

侦查辅助功能通过训练完成的随机森林模型实现，该模型可助力识别犯罪数据中的模式，并提取可解释的决策规则。该框架以随机森林作为核心分类与预测工具，充分利用数据挖掘技术开展犯罪侦查、检测与预测工作。

STCCP在预测准确率上持续优于当前顶尖模型，可提升犯罪防控能力、优化资源配置，并助力构建更安全的社区。