Electronic theses of Computing Science and Mathematics students. http://hdl.handle.net/1893/36 2026-10-28T20:48:28Z 2026-10-28T20:48:28Z Catalano, Giancarlo Antonino Pasquale Ignazio http://hdl.handle.net/1893/38420 2026-10-28T15:53:34Z 2026-06-15T00:00:00Z

Title: Explaining solutions for decision support via Partial Solutions Author(s): Catalano, Giancarlo Antonino Pasquale Ignazio Abstract: Optimisation problems arise across numerous domains, with metaheuristics being widely adopted due to their flexibility and ease of use. Although the metaheuristic framework is relatively simple, the search process is sufficiently complex that it is effectively a black box. Users are often forced to accept results at face value despite the risks of faulty implementations, non-determinism, and unintended biases. To address these limitations, this research introduces an explainability framework for metaheuristics by extracting human-interpretable insights from generated solutions. Central to the proposed methodology is the novel concept of Partial Solutions—defined and implemented as part of this work—which represent understandable sub-structures within metaheuristic solutions associated with high fitness. To find these patterns, we use a multi-objective search method that leverages information relating to their complexity, fitness, and structure. The resulting patterns offer trade-offs between objectives, and each can be post-processed into an explanation. To quantify the “structure” of Partial Solutions, we developed two novel linkage learning methods and a new linkage evaluation procedure, which offers a superior trade-off between accuracy and execution time compared with existing methods. The utility and domain-agnostic nature of this framework are demonstrated and evaluated across several distinct applications. First, the methodology was applied to a complex industrial optimisation problem posed by British Telecom (BT), yielding insightful explanations. Second, the framework was adapted for a recommender system model, successfully generating explanations that satisfy stakeholder objectives. Furthermore, the same underlying methodology was utilised to construct transparent models for predicting solution fitness. To conclude, Partial Solutions are a flexible tool for explainability. They are shown to recover known patterns within benchmark problems, and to be applicable and insightful in large and complex problems such as British Telecom’s Staff Rostering. Our methodology can also be used beyond metaheuristics; for example, we apply it to a recommender system to produce domain-specific explanations. Finally, Partial Solutions can be used to predict fitness: we show that decision trees based on Partial Solutions are competitive with traditional approaches. We conclude the thesis with reflections on whether our objectives have been achieved, and possible future work. Partial Solutions may be used to implement efficient and more understandable metaheuristics, and the applicability of the explanations produced may be expanded beyond combinatorial optimisation.

2026-06-15T00:00:00Z Cui, Zhaoxi http://hdl.handle.net/1893/38355 2026-10-05T10:23:19Z 2025-11-01T00:00:00Z

Title: Network structures and the dynamics of takeover rumours: agent-based simulations of pre-announcement stock price drift Author(s): Cui, Zhaoxi Abstract: Takeover targets often see their stock prices rise sharply before merger or acquisition announcements, mainly because information leaks and related rumours circulate among investors. Although this pattern is well-documented, the underlying process driving such pre-bid increases is still not fully understood. This thesis rebuilds the mechanism from the bottom up: an agent-based market in which an Ignorant–Spreader–Stifler rumour process propagates over a range of networks while heterogeneous value investors and trend followers trade on perceived mispricing and recent returns. Through Monte Carlo simulations, we find a strong link between the magnitude of the run-up and the efficiency of the network topology.

2025-11-01T00:00:00Z Roy Chowdhury, Ipshita http://hdl.handle.net/1893/37730 2026-04-24T14:22:52Z 2025-03-27T00:00:00Z

Title: Understanding Malware using Ontology-based knowledge graph and deep learning techniques Author(s): Roy Chowdhury, Ipshita Abstract: Due to rapid growth and advancement of Big Data and Internet of Things (IoT), industries and organizations are being the main targets of cyber criminals. As IoT enables the linking of all types of `smart' devices industrial appliances, personal information or sensitive data are being constantly attacked by novel malware variants. Therefore, in recent years cyber security has become a major concern due to the emergence of new advanced malware mutants which are causing harm to the users for financial gain to personal information loss. Due to transportation and sharing of huge amount of data, internet-enabled devices and social media platforms are becoming more vulnerable to new generation malware attacks. Because of the limitations of traditional Machine Learning based approaches, it's very difficult to deal with the new malware generations with complex behaviour and varying code structures. Different categories of malware families represents complex, dynamic behaviours and characteristics which can cause a novel and targeted attack in a cyber-system. Existence of large volume of malware types with frequent new additions hinders the cyber resilience effort. This motivation leads to develop a new ontology driven framework that can capture recent malware behaviours and construct novel malware detection and classification framework for new generation malware. This research built an ontology-based knowledge graph to capture metamorphic malware behaviour in the form of API call sequences, and it constructed a dynamic model for implementing an API-based knowledge graph from the domain ontology. The ontological structure are used as the domain knowledge to develop several deep learning algorithms have been incorporated for the prediction of new generation malware classes. In this research a number of deep learning techniques were developed for malware analysis by incorporating CNNs, ResNet-50, GANs, achieving higher accuracies than state-of-the-art models. CNNs and other deep learning architectures were implemented as standalone models and by integrating with other models as well. To address the limitations of static deep learning models, this work includes continual learning approaches to avoid catastrophic forgetting while promoting adaptability in dynamic threat environments. It further addresses two new approaches bringing ontology to Deep Learning: Domain Adaptation, transferring knowledge from one source to one target domain with results pooled in the subsequent stage and then Building Knowledge Graphs using Deep Learning-based Image Classification and Ontology Integration to better support semantic representation through structured integration of the output of classifications into structured knowledge. The overall research work has been carried out to built a meaningful domain specific ontological framework, a wide range of Deep Learning-based neural architectures have been implemented on publicly available benchmark datasets. Employment of various advanced deep learning techniques to improve accuracy and address key challenges such as catastrophic forgetting. Integration of deep learning and ontology with Domain Adaptation and construction of three deep learning based malware knowledge graphs for enhancing interpretability and advancing AI-driven malware analysis in cybersecurity.

2025-03-27T00:00:00Z Shivram, Vivek http://hdl.handle.net/1893/37138 2025-06-17T09:53:48Z 2025-01-01T00:00:00Z

Title: Beyond Intuition: Redefining Internal Audit Risk Assessments Through Machine Learning Paradigms Author(s): Shivram, Vivek Abstract: Established literature has found that Internal Audit Functions (IAF) increasingly view AI- powered innovation as being indispensable to continue staying relevant in the light of in- creased organisational complexity and regulatory expectations. This thesis explores opportunities to deploy Machine Learning (ML) capabilities to assess how IAFs can benefit from increased precision, objectivity and efficiency for delivering internal audit risk assessments. Specifically, this thesis explores how the risk assessment process, a fundamental component of the audit lifecycle, can be redefined through ML to generate measurable value for the IAF. In doing so, this thesis makes several researched assumptions and acknowledges the impact of relevant limitations. Three research questions were framed for evaluation throughout the thesis, to explore the relevance and findings of prevailing literature, identify suitable ML techniques for practical implementation, and measure value from any proposed implementations. In doing so, the thesis supplements the findings of prior academic research, whilst laying the groundwork for further practical improvement by other IAFs. The thesis applied multiple ML techniques to answer these research questions, combining academic ML principles with real-world datasets and implementation techniques, offering a meaningful starting point for IAFs looking to improve their approach to internal audit risk assessments. Three rounds of experiments were performed to reframe risk assessments as specific problem statements for ML models, supported by (1) an examination of the ex- tant risk assessment methodology used by the Senior Leadership Team (SLT) at the subject organisation, a world-leading financial services firm, (2) an analysis of organisational data relevant to the risk assessment, anonymised for the purposes of this thesis, and (3) the construction and application of several ML hypotheses on this dataset. The conclusions of these experiments were consolidated to answer the research questions, including an assessment of how the value arising from such efforts could be measured for continued sponsorship and improvement. Overall, the thesis noted that there is no single best ML approach for practical implementation in all cases, and that the choice of ML models would vary based on the quality of underlying data, consistent audit methodologies, as well as the size of the dataset including the final set of input features. The thesis found that a supervised machine learning approach was most suitable for internal audit risk assessments, given the high class-specific precision ratio of the final model (c.88%) on the test dataset. However, the thesis noted significant dependencies on high-quality datasets and a requirement to carefully curate relevant features to ensure meaningful model outputs. Although the thesis evaluated unsupervised machine learning techniques, these were discounted from practical implementation due to comparatively lower accuracy, inhibiting practical implementation for the given dataset. However, the thesis also noted that such techniques may generate better results in alternative practical settings, where significantly larger datasets are used for risk assessment cycles.

2025-01-01T00:00:00Z