https://dl.ucsc.cmb.ac.lk/jspui/handle/123456789/4995 Sat, 18 Jul 2026 15:15:39 GMT 2026-07-18T15:15:39Z https://dl.ucsc.cmb.ac.lk/jspui/handle/123456789/5000 Title: Unveiling Patterns in Online English Business News in Sri Lanka: Abstractive Text Summarization with Large Language Models Authors: Wimalarathna, V.M. Abstract: ABSTRACT The emergence of online business news has amplified the rapid generation of large volumes of textual data, making it an important resource for businesses to extract business insights to support business decision making. The aim of this research project is to propose an approach to leverage online English business news to gain business insights with specific focus on Sri Lanka. Accordingly, the research employs a combined approach of document classification, document clustering, and abstractive text summarization methods to utilize large amount online business news data to gain meaningful insights. The research methodology incorporates Bidirectional Encoder Representations from Transformers (BERT) model for news classification, followed by K-means news clustering, and cluster-wise abstractive text summarization utilizing large language models (LLMs), including Llama3, Mistral, and DeepSeek R1 along with retrieval-augmented generation (RAG) to enhance the accuracy and relevance of the generated content. The dataset is obtained from Sri Lankan business news sources, including DailyFT and Lanka Business Online. The ability to fine-tune BERT for specific classification tasks, combined with the scalability of Kmeans clustering and the text generation capabilities of LLMs accommodates unveil business insights from large volumes of textual data. However, there is still room for improvement by experimenting with large datasets and improving above techniques to enhance the performance of proposed approach. The results of the study indicate that the combination of machine learning methods, such as classification and clustering, along with LLMs, yields better results by complementing each method’s strengths, to leverage textual data more efficiently than applying them in isolation. This makes the proposed approach is more pragmatic way of leveraging computational method to handle large amount of online business news. Therefore, current study lays the foundation for harnessing the capabilities of machine learning and LLMs for real-world business applications, specifically in developing tools that support decision-making in business by utilizing online business news data Sun, 15 Jun 2025 00:00:00 GMT https://dl.ucsc.cmb.ac.lk/jspui/handle/123456789/5000 2025-06-15T00:00:00Z https://dl.ucsc.cmb.ac.lk/jspui/handle/123456789/4999 Title: Mitigating Risk of Bank Loans by Predicting Maximum Loan Amount and Approval using Machine Learning Authors: Senanayaka, S. K. D. V. L. Abstract: ABSTRACT In current financial system banks play major role, they are engaged in provision of liquidity to the entire economy. Bank loan represent the significant source for these bank and financial institutions and loan approval process plays, a critical role in economic stability and growth. Accurate bank loan prediction is crucial for financial institutions to assess customer creditworthiness and minimized the default risk. Rejecting a loan request without providing the maximum approvable amount can lead to customer loss. The aim of this study is to mitigate the risk of loan approval process in a bank by training machine learning models to predict the maximum loan amount and supporting the decision making in the loan approval process. Datasets (two) are download from Kaggle and combined. The research applies classification models - Decision Tree, Random Forest, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Logistic Regression, and Artificial Neural Networks to predict the loan approval. Also regression models - Linear Regression, Random Forest Regression, Gradient Boosting Regression, Support Vector Machine (SVM), and Decision Tree Regression which used to predict the maximum loan amount. The models are evaluated using 5-flod cross validation, whit classification performance access using accuracy, precision, recall, F1- score, confusion matrix, and ROC curve. While regression models are analyzed using Mean Absolute Error (MAE) and Mean Squared Error (MSE). The Random Forest model is the best performer in loan approval process with accuracy of 95%, The Random Forest Regression model is the best performer in maximum loan amount prediction model with minimum MSA (0.1272) & MAE (0.1558) values. Based on evaluation results Random Forest classifier selected as suitable model for bank loan approval prediction which have best performance compared with other models. Sun, 22 Jun 2025 00:00:00 GMT https://dl.ucsc.cmb.ac.lk/jspui/handle/123456789/4999 2025-06-22T00:00:00Z https://dl.ucsc.cmb.ac.lk/jspui/handle/123456789/4998 Title: Anomaly Detection in Equity Stocks: Applying Machine Learning Techniques to Trading Data from the Colombo Stock Exchange Authors: Jayasinghe, J.M.G.D Abstract: ABSTRACT Anomaly Detection in Equity Stocks: Applying Machine Learning Techniques to Trading Data from the Colombo Stock Exchange explores the possibility of adopting machine learning models to identify abnormal trading behaviors that could indicate market manipulation. Motivated by the need to enhance investor confidence by researching methods that can detect market abnormalities, this research leverages publicly available historical market data from the Colombo Stock Exchange. The study adopted two approaches: a baseline time-series forecasting method to detect anomalies, named Facebook Prophet. A series of secondary models had Isolation Forest as a single model, and as hybrid models, Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), and autoregressive LSTM (AR-LSTM) forecasting models were incorporated with Isolation Forest as the unsupervised anomaly detection method. The research has considered Commercial Leasing and Finance PLC (COLE) as the main focus due to the SEC Sri Lanka confirming manipulations from 9th to 25th of August 2021, along with John Keells PLC (JKH) as a control dataset, which does not have SEC Sri Lanka confirmed manipulation cases. In terms of COLE, the Prophet model achieved the highest F1-Score of 0.1792, balancing precision (0.1111) and recall (0.4615), making Prophet the most effective overall. AR-LSTM + Isolation Forest demonstrated the highest recall of 0.6154, but at the cost of precision 0.0816, resulting in a slightly lower F1-Score of 0.1440. Isolation Forest and CNN+ Isolation Forest showed comparable output, while LSTM + Isolation Forest underperformed with the lowest F1-Score of 0.0740. For JKH, using a baseline assumption that all data points were anomalous, Prophet again emerged as the leading model with an F1-Score of 0.1110, followed by AR-LSTM and LSTM (0.094 each), validating Prophet’s robustness across both verified and unverified anomaly detection contexts. The findings of this study suggest that anomaly detection through machine learning models by solely relying on publicly available data, such as daily closing prices, trading volumes, along with derived technical indicators, is an effective approach that will provide a valuable early-warning signal to market participants. Future work can incorporate expanded features along with sentiment data to enhance anomaly detection using machine learning models. Keywords: Anomaly Detection, Machine Learning, Facebook Prophet, LSTM, Isolation Forest, CNN, AR-LSTM, Stock Market Manipulation, Colombo Stock Exchange, Time-Series Forecasting. Mon, 23 Jun 2025 00:00:00 GMT https://dl.ucsc.cmb.ac.lk/jspui/handle/123456789/4998 2025-06-23T00:00:00Z https://dl.ucsc.cmb.ac.lk/jspui/handle/123456789/4997 Title: New Evaluation Metric for Large Language Models in Summarization Authors: Dilshan, D.K.W.S Abstract: ABSTRACT This thesis introduces and evaluates a new text summarization evaluation metric, SynSemScore, to fill the gap between the simple lexical-based metrics such as BLEU, ROUGE, and METEOR, and the complex embedding-based methods. While the previous methods are based mainly on surface level word matching or are costly in terms of computational complexity, SynSemScore lies in between by using syntactic matching together with semantic similarity measured using WordNet-like structures and a redundancy penalty to identify redundant content. The purpose is to offer a simple but effective way of measuring the quality of the machine produced summaries in terms of coherence, consistency, fluency and relevance. To validate SynSemScore empirically, this research employs the SummEval dataset which offers a variety of human annotated summaries for CNN/Daily Mail articles. The data is cleaned up thoroughly to ensure data integrity and correlation analyses are made between SynSemScore and other metrics. The results of the experiments reveal that SynSemScore is more relevant to the human rating, outperforming traditional approaches in several aspects. Hence, the proposed metric can be considered as a feasible solution for assessing the quality of summarization models on a large scale without the need for expensive training of transformer-based models. Besides the technical aspects of SynSemScore, the work describes the complete process of the research, including data collection and preparation, hyperparameter optimization, and correlation analysis of the metric. Finally, SynSemSemScore presents a flexible, easy-to-interpret, and semantic-aware method for assessing the performance of current summarization systems and can therefore help to enhance the quality of the generated text in both academic and business applications. Keywords: Text Summarization, SynSemScore, Semantic Similarity, Redundancy Penalty, SummEval, Evaluation Metrics, Large Language Models. Thu, 12 Jun 2025 00:00:00 GMT https://dl.ucsc.cmb.ac.lk/jspui/handle/123456789/4997 2025-06-12T00:00:00Z