https://www.um.edu.mt/library/oar/handle/123456789/10779 Thu, 11 Jun 2026 17:59:26 GMT 2026-06-11T17:59:26Z https://www.um.edu.mt/library/oar/handle/123456789/146931 Title: Leveraging invariant prediction for mitigating specificity constraints in affect modelling Abstract: Affect modelling aims to predict human emotional states from multimodal signals, yet current approaches often struggle to generalise beyond the specific datasets or contexts in which they are trained. This dissertation investigates the use of in variant features, predictors whose relationship with affective states remains stable across distinct environments, as a strategy to improve generalisability. To this end, two publicly available corpora, AGAIN and RECOLA, were systematically parti tioned into environments defined by user, task, and annotator triplets. An envi ronment refers to the conditions under which data is collected, and data gathered within the same environment is assumed to come from the same underlying distri bution. The Invariant Causal Prediction (ICP) framework was employed to identify stable features across these environments, which were then compared against full feature sets and principal components derived through PCA. Three supervised learning models—Logistic Regression, a feed-forward Neural Network,and a Long Short-Term Memory (LSTM)network — were trained under all three feature conditions, using group-based cross-validation to avoid information leakage. Results demonstrate that invariant features can deliver measurable benefits for feed-forward models, particularly in enhancing accuracy and correlation while substantially reducing feature dimensionality. However, their advantages were less consistent for sequence models like LSTMs, where temporal dependencies were not fully captured by invariants alone. Statistical significance tests further showed that invariant features improved balanced classification (F1) more strongly in AGAIN than in RECOLA,underscoring the dataset-specific nature of their effectiveness. Overall, the findings highlight both the promise and the limitations of invariance in affect modelling. While not a universal solution, invariant features represent a principled means of isolating robust predictors across heterogeneous contexts, contributing to the broader goal of developing affective systems that are reliable, interpretable, and adaptable across diverse real-world settings. Description: M.Sc.(Melit.) Wed, 01 Jan 2025 00:00:00 GMT https://www.um.edu.mt/library/oar/handle/123456789/146931 2025-01-01T00:00:00Z https://www.um.edu.mt/library/oar/handle/123456789/146930 Title: Siamese network‐based vector embeddings of MRI scans for twin identification Abstract: Monozygotic twins are identical twins that develop from a single fertilised egg that spontaneously splits, resulting in two individuals sharing 100% genetic material. Identifying monozygotic twins from brain MRI scans represents a frontier challenge in computational medical imaging with significant implications for understanding genetic influences on neuroanatomical structure through direct pattern recognition. While classical twin studies using ACE models decompose statistical variance to establish independent regional heritability estimates (60‐80%), this study introduces a fundamentally different computational framework that learns directly from MRI data to rank neuroanatomical regions by their collective discriminative capacity for genetic similarity detection, complementing traditional statistical approaches through data‐driven analysis. Adeep learning methodology employing Siamese networks with 3D CNN backbones is developed for automated twin identification using 138 genetically verified monozygotic twin pairs (276 subjects) from the Human Connectome Project S1200 dataset. Modified U‐Net, ResNet, and DenseNet architectures generate 128‐dimensional embeddings optimised via triplet loss with hard negative mining, forcing models to learn subtle genetic signatures by focusing on challenging discriminative examples that distinguish twins from their most similar morphological matches. U‐Net achieved superior computational performance with 92.0% F1‐score (σ = 2.5%), 95.2% AUC‐ROC, and 91.4% accuracy, while ResNet demonstrated competitive results (89.6% F1‐score) and DenseNet showed greater variability (88.5% F1‐score). Embedding analysis reveals clear bimodal separation between genetically related and unrelated individuals through learned morphological patterns. Layer‐Wise Relevance Propagation analysis provides the first data‐driven ranking of neuroanatomical regions by discriminative importance for genetic relatedness detection. Statistical analysis reveals pronounced subcortical dominance with large effect size (Cohen’s d = 2.80, p = 3.89e‐6), with six subcortical structures occupying top positions, including the thalamus (0.955), brainstem (0.875), and hypothalamus (0.707). This computational hierarchy contrasts with traditional ACE studies reporting highest heritability in cortical areas (frontal 78‐95%, temporal 77‐89%), demonstrating that direct pattern recognition from MRI data identifies different neuroanatomical signatures than statistical variance decomposition. Notably, models utilise practically all brain regions (most importance scores > 0.2), indicating distributed multivariate processing rather than selective regional dependence. Ablation studies confirm data augmentation’s critical role, with substantial i performance improvements across CNN architectures. Clinical integration through standard neuroimaging formats in Connectome Workbench demonstrates immediate practical utility, positioning this computational approach for adoption in research and clinical environments requiring direct analysis of genetic influences in brain structure. The framework advances precision neuroimaging by providing automated, quantitative genetic similarity detection through direct pattern recognition, revealing spatial insights that complement traditional heritability studies while offering methodological advances applicable to diverse medical imaging classification tasks requiring regional discriminative analysis Description: M.Sc.(Melit.) Wed, 01 Jan 2025 00:00:00 GMT https://www.um.edu.mt/library/oar/handle/123456789/146930 2025-01-01T00:00:00Z https://www.um.edu.mt/library/oar/handle/123456789/146924 Title: Advancing automated Maltese spell checking using deep learning Abstract: Research in Natural Language Processing (NLP) for Maltese has advanced in recent years, yet reliable and contextually aware spell checking support remains limited. Existing tools, such as spelling.mt, provide dictionary-based correction but cannot handle grammatical or context-sensitive errors, leaving a clear gap com pared with high-resource languages. This dissertation investigates whether a Large Language Model (LLM) can be fine-tuned to develop an effective and resource efficient spell checker for Maltese. The study adapts Meta’s Large Language Model MetaAI (LLaMA)-3-8B-Instruct model using pairs of incorrect and corrected Maltese words and sentences. A cus tom fine-tuning corpus was created by introducing linguistically motivated syn thetic errors into correctly written text and incorporating smaller sets of authen tic incorrect-correct pairs from prior research. Two complementary models were trained: a word-level model that corrects individual tokens and a sentence-level modelthatcorrectsgrammaticalandorthographicerrorswithinfullsentences. Their performance wasevaluated across synthetic, real-world, and correct-input datasets. The main contribution of this dissertation is the development of the first LLM based spell checking models, specifically fine-tuned for Maltese, supported by a curated error-correction dataset and a systematic multi-setting evaluation. To our knowledge, this study is the first to explore the use of a modern generative LLM to build a Maltese spell checker and assess its potential in a low-resource setting. The developed models demonstrated strong performance on incorrect-correct pairs with synthetically generated errors, with the word-level model achieving 84.64% accuracy and the sentence-level model reaching 95.3%. For real-world eval uation, only the sentence-level model was tested because no suitable word-level datasets are available. In this context, the accuracy of the sentence-level model dropped to 29.6%, reflecting the increased variability of naturally occurring mis takes. However, metrics such as Error Annotation Toolkit (ERRANT), Bilingual Evaluation Understudy (BLEU), and Grammar Language Evaluation Understudy (GLEU) returned promising values, indicating that the model often produced cor rections close to the target. Overall, the findings demonstrate that fine-tuning an LLM offers a promising pathway towards effective Maltese spell checking, while revealing key challenges related to generalisation and data scarcity. The study provides an initial bench mark andoutlines concrete directions for developing more robust, deployable spell checking solutions for low-resource languages. Description: M.Sc.(Melit.) Thu, 01 Jan 2026 00:00:00 GMT https://www.um.edu.mt/library/oar/handle/123456789/146924 2026-01-01T00:00:00Z https://www.um.edu.mt/library/oar/handle/123456789/146923 Title: Enhancing transparency and interpretability in AI-driven algorithmic trading Abstract: Algorithmic trading increasingly relies on AI-driven decision systems, yet opaque models limit trust and accountability. This study investigates how Reinforcement Learning (RL) can be made more transparent by combining a model-agnostic explainability framework with Large Language Model (LLM) based narrative synthesis. The framework comprises four layers linking trading behaviour to feature attribution and temporal dynamics, stability and regime sensitivity, policy surrogacy, and reward decomposition. In Experiment 1, we apply state-of-the-art RL algorithms to constituents of the Dow Jones Industrial Average and evaluate performance using standard return and risk measures. We investigate modern explainability techniques across market regimes to characterise which indicators drive decisions, how stable attributions are over time, and how policies can be approximated by compact surrogate rules. The results indicate convergent feature drivers, expected masking behaviour, smoother attributions, indicating that explanations are temporally stable, and low-complexity surrogates with credible fidelity. In Experiment 2 we extend these findings by investigating how explainability artefacts can be translated into grounded natural-language narratives. Explanations generated from the framework are assessed through automated text metrics and a human-centred study with participants at varying levels of trading experience. Our results show that structured prompting improves lexical quality and adherence to factuality constraints relative to a zero-shot baseline. An important finding is that participants view the combined visual–narrative explanations as clear, moderately trust worthy, and practically helpful, with narratives particularly useful for reconstructing the agents’ reasoning and feature contributions. The contributions of this work are threefold. First, the study introduces a unified, model-agnostic explainability framework for financial RL linking feature-level attributions to policy behaviour and realised rewards. Second, it proposes and validates a pipeline for grounded natural-language synthesis anchored in quantitative explainability outputs. Third, in daily DJIA trading and a small human study, it provides evidence that technical faithfulness and human-centred accessibility can be advanced together, offering a template for transparent decision support in finance. Description: M.Sc.(Melit.) Thu, 01 Jan 2026 00:00:00 GMT https://www.um.edu.mt/library/oar/handle/123456789/146923 2026-01-01T00:00:00Z