Real Time Crime Prediction Using Social Media

There is no doubt that crime is on the increase and has a detrimental influence on a nation's economy despite several attempts of studies on crime prediction to minimise crime rates. Historically, data mining techniques for crime prediction models often rely on historical information and its mostly country specific. In fact, only a few of the earlier studies on crime prediction follow standard data mining procedure. Hence, considering the current worldwide crime trend in which criminals routinely publish their criminal intent on social media and ask others to see and/or engage in different crimes, an alternative, and more dynamic strategy is needed. The goal of this research is to improve the performance of crime prediction models. Thus, this thesis explores the potential of using information on social media (Twitter) for crime prediction in combination with historical crime data. It also figures out, using data mining techniques, the most relevant feature engineering needed for United Kingdom dataset which could improve crime prediction model performance. Additionally, this study presents a function that could be used by every state in the United Kingdom for data cleansing, pre-processing and feature engineering. A shinny App was also use to display the tweets sentiment trends to prevent crime in near-real time.Exploratory analysis is essential for revealing the necessary data pre-processing and feature engineering needed prior to feeding the data into the machine learning model for efficient result. Based on earlier documented studies available, this is the first research to do a full exploratory analysis of historical British crime statistics using stop and search historical dataset. Also, based on the findings from the exploratory study, an algorithm was created to clean the data, and prepare it for further analysis and model creation. This is an enormous success because it provides a perfect dataset for future research, particularly for non-experts to utilise in constructing models to forecast crime or conducting investigations in around 32 police districts of the United Kingdom.Moreover, this study is the first study to present a complete collection of geo-spatial parameters for training a crime prediction model by combining demographic data from the same source in the United Kingdom with hourly sentiment polarity that was not restricted to Twitter keyword search. Six unique base models that were frequently mentioned in the previous literature was selected and used to train stop-and-search historical crime dataset and evaluated on test data and finally validated with dataset from London and Kent crime datasets.Two different datasets were created from twitter and historical data (historical crime data with twitter sentiment score and historical data without twitter sentiment score). Six of the most prevalent machine learning classifiers (Random Forest, Decision Tree, K-nearest model, support vector machine, neural network and naïve bayes) were trained and tested on these datasets. Additionally, hyperparameters of each of the six models developed were tweaked using random grid search. Voting classifiers and logistic regression stacked ensemble of different models were also trained and tested on the same datasets to enhance the individual model performance.In addition, two combinations of stack ensembles of multiple models were constructed to enhance and choose the most suitable models for crime prediction, and based on their performance, the appropriate prediction model for the UK dataset would be selected. In terms of how the research may be interpreted, it differs from most earlier studies that employed Twitter data in that several methodologies were used to show how each attribute contributed to the construction of the model, and the findings were discussed and interpreted in the context of the study. Further, a shiny app visualisation tool was designed to display the tweets’ sentiment score, the text, the users’ screen name, and the tweets’ vicinity which allows the investigation of any criminal actions in near-real time. The evaluation of the models revealed that Random Forest, Decision Tree, and K nearest neighbour outperformed other models. However, decision trees and Random Forests perform better consistently when evaluated on test data