Browsing by Author "Lakshan, W. S. V."

Now showing 1 - 3 of 3

Comparison between the Machine Learning Algorithms to determine the suitable input features for personal theft, sexual assault, and house burglary victimization prediction
(Faculty of Science, University of Kelaniya Sri Lanka, 2022) Lakshan, W. S. V.; Weerakoon, W. A. C.
The crime-related predictions can be vastly supported by most of the available supervised machine learning models. The possibility of becoming a victim increases daily in each crime category. The main difficulty is to find how severe the impact is upon the victim after the crime. Here, the Random Forest, Support Vector Machine (SVM), K-Nearest Neighbour (KNN) algorithm, and Neural Network models were compared with the use of available features found from a secondary dataset to build a better prediction model, which has been implemented in four main phases over two aspects based on the possibility of becoming a victim and severity of the crime. The available features were used as the inputs for phase I and Principal Component Analysis and correlation tests were performed to identify the appropriate and essential feature combinations for the rest of the phases. The pre-processed datasets were used to implement and train the models. Moreover, the Random Forest model was proven to be the most efficient model with an accuracy of 85.33% in phase four when comparing the accuracy levels of the models over different phases, while the KNN and Neural Network models obtained an accuracy of over 70% and SVM obtained the least accuracy in the same phase. In phase one, the Random Forest algorithm was executed with a precision of 76%, while KNN and Neural Network model obtained around 70%. The final outputs obtained for phase four showed that factors such as age, year, gender, race, and relationship with the perpetrator will be the most suitable features to build an accurate machine learning model for victimization prediction. The mentality level of the offender and intention of doing it has the main impact on the severity level. Also, authorities need to keep track of the fact whether it is a repeat offence or not, the main offender or not and the contribution of the offender to support better information inputs for the prediction models. This study developed a victimization prediction model with reference to personal theft, sexual assault, and house burglary. This would be a step forward from previous research works of rule-based victimization possibility index prediction for small victim clusters. Further, new features were identified in the last phase, which can be used to develop models to predict criminal behaviour after sending them back to the society. This will greatly benefit the authorized bodies to monitor them and reduce the possibility of victimization.
An efficient credit score prediction model with Deep Neural Network
(Faculty of Science, University of Kelaniya Sri Lanka, 2022) Lakshan, W. S. V.; Katugampala, K. D. V. C.
The rapid expansion of the financial industry over the evolution of Computer Science is a breakthrough for both fields. Financial institutes tend to get the maximum benefit from computer science by analysing huge stack of unstructured data which increases daily. One of the most common facilities those institutes provide is loan facilities for both individuals and businesses. Before taking a decision, they need to consider many factors and a wide variety of reference documents to lend on loan. The credit score of a person or a business is an important aspect that needs to be considered when lending money. It depicts the ability of the borrower to pay back the loan on time. This research focused on developing a credit score prediction model based on a deep neural network to predict creditworthiness. Two separate models had been created for personal loans and micro loans. In the initial phase, data had been pre-processed, and correlation tests were carried out for the input feature selection against the prediction of approval. We performed the dimensionality reduction based on principal component analysis to discard components that have low information related to the credit score. This study will support machine learning algorithms to explore, analyse and visualize the data with a more efficient approach. The deep neural network model was trained with the pre-processed data and tuned for the best model by changing weights and activation functions. A REST API was developed as a plugin using the model which is to be integrated into prevailing systems of institutions. Therefore, the overall architecture goals of the system were to provide a high-functioning REST API with a low response time (150 – 200 ms), and to predict the creditworthiness of a client with the details of the relevant inputs which have been achieved with an 80% level of accuracy for the personal loan approval model without tuning the hyperparameters and 72.6% level of accuracy micro loan approval model with the hyperparameter tuning. We achieved better prediction performance of the models by adjusting the hyperparameters such as weight values and biases of the Recurrent Neural Network model selected for model building as the deep learning algorithm. To make the prediction model more accurate and precise, it is required to thoroughly identify the impact points and key features of the relationship between the financial institutes and the borrower.
An enhanced ensemble model for crime occurrence prediction using machine learning
(Faculty of Science, University of Kelaniya, Sri Lanka., 2021) Lakshan, W. S. V.; Silva, A. T. P.; Weerakoon, W. A. C.
With the rapid increase of crime, law enforcement departments are struggling to stop crimes and continuously demand automated advanced systems for crime control to provide better protection to the human being in a community. Crime predication plays a vital role in crime control. Crime analysis & prediction can reveal the complexities and hidden patterns in the crime datasets, and it can be used for early decision making. The early researchers attempted to predict the crime using a machine learning model with main factors including time, date and location but overlooked other essential factors. This paper aims to present an enhanced crime prediction algorithm based on ensemble classification technique while identifying several factors that affect the learning model's performance. The correlation of the factors versus the prediction label is analyzed using the Spearman and Pearson techniques to determine the important, influential factors. The prediction model was developed based on Ensemble techniques using the Random forest model with the Voting Classifier. Multiple decision trees had implemented the crime prediction model of this research as the base model and the Logistic Regression and K-Nearest Neighbor algorithm as the sub-models. The final classifier was developed based on using the Graphical User Interface and the REST API methods to predict the possibilities of the crime occurrences at a given specific time in each location. The proposed method can identify the likelihood of a crime in a particular location at a specific time. This helps to implement better strategic and tactical ways to minimize crimes with less risk, as the accuracy of the crime prediction algorithm was 89%.