http://dspace.dtu.ac.in:8080/jspui/handle/123456789/51 2026-07-01T06:01:15Z 2026-07-01T06:01:15Z KAMBOJ, PRADEEP KUMAR, SHAILENDER ( SUPERVISOR) GOYAL, VIKRAM (CO - SUPERVISOR ) http://dspace.dtu.ac.in:8080/jspui/handle/repository/22944 2026-06-25T05:08:43Z 2026-04-01T00:00:00Z

Title: SOCIAL BIAS IDENTIFICATION AND MITIGATION IN NATURAL LANGUAGE TEXT USING MACHINE LEARNING Authors: KAMBOJ, PRADEEP; KUMAR, SHAILENDER ( SUPERVISOR); GOYAL, VIKRAM (CO - SUPERVISOR ) Abstract: Advanced Artificial Intelligence (AI) methods have enabled the creation of sophisticated large language models (LLMs) capable of generating human-like text and handling a broad spectrum of complex language comprehension tasks. The last decade has seen the advent of LLMs that fill crucial roles across a variety of applications, including automated content generation and summarization, healthcare analytics, legal decision support, conversational agents, and educational technologies. Despite their remarkable abilities, these models often reflect and even amplify the social biases embedded in the large datasets on which they are trained. These biases can manifest as stereotypes or unjust associations related to gender, race, religion, profession, or other social features. When these AI systems are deployed in high-stakes domains where fairness and reliability are paramount, the presence of such biases raises major ethical, social, and technical concerns. As a result, understanding, measuring, and mitigating bias in LLMs has emerged as a prominent research challenge at the forefront of responsible and trustworthy AI. This thesis constitutes a thorough exploration of social bias in natural language text generated by language models (LMs) and LLMs, with a focus on systematic approaches to measuring, evaluating, and mitigating it. The research draws on theoretical, empirical, experimental, and methodological approaches to investigate bias from several angles across the AI pipeline, including word embeddings, contextualized language models, prompt-based inference functions, and fine-tuning strategies. The work focuses on understanding biases across these components and seeks practical solutions to build fairer and more trustworthy generative AI systems. The initial phase of the research investigates gender bias in contextualized word embeddings generated by transformer-based LMs. Word embeddings are the building blocks of language in many NLP systems, and biases encoded in these representations can carry over to downstream applications. The gender direction in the embedding space is extracted, and the gender polarity of profession-related terms (occupation names) with respect to gendered pronouns is calculated, yielding a quantitative framework for measuring one type of bias: that women or men are less likely to pursue certain professions. Indeed, an experimental analysis shows that dynamic embeddings from transformer-based models exhibit substantial gender associations even in the absence of explicit gender information in the input text. To alleviate this problem, we propose a form of post-processing debiasing that modifies the embedding representations to reduce stereotypical associations while preserving the semantic relationships among words. The experimental results show that the proposed method can significantly alleviate gender bias in profession embeddings, thereby balancing the model’s representations. Building on this foundation, the thesis broadens the analysis to large language models and a wider range of societal biases stemming from multiple demographic attributes. We introduce a systematic evaluation framework for bias in LLM-generated outputs, in part by creating a curated inference dataset from previously established bias benchmarks. The dataset includes contexts that encourage language models to generate stereotypical, anti-stereotypical, and neutral responses, enabling systematic assessment of model behaviour. This study provides a comprehensive mechanism for v analyzing how different models respond to socially sensitive contexts and how bias manifests in generated text. This research makes an important contribution by exploring prompt engineering to both detect and mitigate bias in LLMs. Several types of prompt variants are developed to investigate the effects of their design on model behaviour, namely standard, chain of-thought, cognitive-style, and human-persona prompts. These prompts are systematically assessed to study the effects of various prompting techniques on output bias. Also proposed are the debiased versions of these prompts that explicitly elicit neutral reasoning and unbiased decision-making. The introduction of prompt-only bias evaluation is a key aspect of the extended work, exploring whether biased responses can be induced by prompts alone, without context. Experimental results indicate that when certain prompts are presented to language models, those models make stereotypical predictions, suggesting that bias arises from the interaction between prompts and the models' reasoning mechanisms, rather than solely from the training data. This underlined the importance of careful prompt design and evaluation when deploying language models in real-world settings. Alongside this bias analysis, the research also delves into the issue of hallucination in LLMs, whereby a model provides confident answers that are factually incorrect or unsupported. Across most domains, hallucinations undermine the model’s reliability and may introduce risks in critical domains such as healthcare, legal advice, and policy analysis. To tackle this phenomenon, the thesis presents a contrastive decoding method powered by disturb prompts to compare the probability distributions of model outputs for same prompt and perturbation-prompt scenarios. The method helps detect hallucinated content and enhances the factual consistency of outputs by comparing responses to normal prompts with those to perturbed prompts. The results show that contrastive prompting methods can mitigate hallucination and improve the robustness of language model outputs. Another important aspect of the research is assessing how well fine-tuning approaches mitigate biases. Among such models, large open-source language models are fine tuned on balanced sets with equal numbers of biased/unbiased statements across a wide range of social categories. Fine-tuning is when models are trained to produce more neutral and fair responses while retaining their language comprehension. In fact, experimental results show that fine-tuning with fairness-aware special prompts significantly reduces the model's biased outputs and improves fairness performance. In conclusion, the work in this thesis demonstrates that bias in LMs is a complex, multifaceted phenomenon with multiple underlying sources, including training data, representation learning, and prompting. Tackling this challenge requires the integrated use of bias measurement, dataset design, prompt engineering, model fine-tuning, and evaluation metrics. The methodologies are cross-disciplinary, offering actionable tools to identify and prevent bias in generative AI systems without sacrificing performance or usability. This work extends beyond technical contributions, establishing the need for a broader meaning of fair and responsible development in the internalization of AI. Overall, this thesis gives a good overview of bias in LMs and LLMs. The research, by integrating representation-level analysis, prompt-based evaluation, hallucination detection, and fairness-aware fine-tuning, provides novel insights into the mechanisms that produce vi biases in AI systems while suggesting appropriate strategies to mitigate them. The results of this work demonstrate the potential to help establish more ethical, fair, transparent, and socially responsible generative AI technologies that can serve a wider range of communities without perpetuating harmful stereotypes or obesity-related inequalities.

2026-04-01T00:00:00Z AHUJA, CHIRAG http://dspace.dtu.ac.in:8080/jspui/handle/repository/22764 2026-06-08T05:46:01Z 2026-02-01T00:00:00Z

Title: EEG SIGNAL CLASSIFICATION USING FEW-SHOT LEARNING Authors: AHUJA, CHIRAG Abstract: Electroencephalogram (EEG) signals are crucial in various applications, including Motor Imagery, Emotion Recognition, Visual Evoked Potentials, and Mental Workload assessment. However, EEG classification remains challenging due to limited labelled data, high noise levels, and substantial inter- and intra-subject variability. This thesis addresses these challenges by leveraging Few-Shot Learning (FSL) techniques to enable e!ective learning from minimal data for EEG signal classification. To overcome key limitations, this research integrates Data Augmentation, Transfer Learning, and Self-Supervised Learning (SSL) within the FSL framework. Specifically, it focuses on (1) developing EEG-specific data augmentation strategies to mitigate data scarcity, (2) designing transfer learning methodology to facilitate e”cient knowledge transfer across subjects, and (3) formulating SSL methods to enhance FSL with minimal labelled data. Firstly, the thesis presents a comprehensive literature review of FSL techniques in EEG classification, detailing data augmentation, transfer learning, and SSL methodologies. It establishes best practices for FSL for EEG classification and provides standardized guidelines for reporting results in future studies. Secondly, it explores data augmentation techniques to reduce dependence on limited EEG datasets by generating realistic augmented samples. It introduces Auto- Augmentation for Emotion Recognition in EEG - A Class and Subject Invariant Approach (ADAPTER) framework, which, when integrated with the cross-subject model Self-Organizing Graph Neural Network (SOGNN), achieves around 2% F1- score gain over vanilla SOGNN achieving 88.54% of cross-subject accuracy on SEED. Thirdly, recognizing the need for improved subject adaptation, the thesis proposes a novel framework called Transfer and Robust Adaptation of New Subjects in EEG vi Technology (TRANSIT-EEG). It combines a subject-specific data-augmentation - Individualised Denoising Probabilistic Model (IDPM) with Low-Rank Adaptation (LoRA) based transfer learning on an enhanced SOGNN model called Self-Organizing Graph Attention Transformer (SOGAT). Experimental evaluations on SEED and Phyaat datasets demonstrate superior cross-subject F1 scores of 91.53% and 87.78%, respectively. Finally, the work addresses cross-device generalization in EEG classification through two Self-Supervised Learning frameworks: (i) Self-Supervised Enhancement for Multidimensional Emotion Recognition using GNNs for EEG (SS-EMERGE) and (ii) Unified Framework for Yielding EEG-based Emotion Recognition Model with Self-Supervised Learning (UNIFY-ESSL). SS-EMERGE employs a multidimensional architecture to capture temporal, spectral, and spatial features. A meiosis-based data-augmentation pretext task drives cross-subject generalization. The model delivers Macro-F1 scores of 92.35% and 81.51% on SEED and SEED-IV, respectively. When fine-tuned with only half of the labels, it still achieves 86.13% and 76.75% on SEED and SEED-IV, respectively. UNIFY-ESSL evaluates Contrastive Learning (SimCLR) and Contrastive Predictive Coding (CPC) based pretext tasks alongside a proposed data sampling strategy. The experimental results show that SimCLR attains F1- scores of 82.62%, 87.83%, and 89.05% on SEED, DEAP, and DREAMER datasets, respectively, while CPC achieves 81.35%, 82.27%, and 91.23%. It improves cross- dataset generalization, with a 1-2% performance gain on DREAMER and maintained performance on DEAP despite channel reduction, although SEED experiences a 3% F1-score drop due to significant channel reduction. These contributions enable realistic data augmentation, rapid adaptation to new subjects for personalization, and unified modeling across datasets—advancing robust, adaptable, and generalizable EEG classification for diverse real-world applications.

2026-02-01T00:00:00Z ZIYA, FATIMA Kumar, Sanjay (SUPERVISOR) http://dspace.dtu.ac.in:8080/jspui/handle/repository/22763 2026-06-08T05:45:54Z 2026-02-01T00:00:00Z

Title: DEVELOPMENT OF LINK PREDICTION MODEL IN SOCIAL NETWORK Authors: ZIYA, FATIMA; Kumar, Sanjay (SUPERVISOR) Abstract: Link prediction in social networks plays a crucial role in understanding network evolution, identifying potential interactions, and supporting applications such as rec- ommendation systems, community analysis, and the discovery of biological networks. The fundamental problem of link prediction is to estimate the likelihood of future or missing connections between pairs of nodes based on existing network information, structural patterns, node attributes, and temporal evolution. However, real-world net- works are highly complex, sparse, dynamic, and heterogeneous, making traditional similarity-based and shallow learning approaches insufficient to capture deep struc- tural semantics and evolving behavioral patterns. In this thesis, we introduce a robust and adaptive approach to link prediction in social networks. The present study integrates traditional similarity-based techniques with advanced deep music recommendations, among effective similarity scores ex- isting methods for list structure- and attribute-aware information, a single similarity index, or paths from performance and reliability of the proposed methodology. The first model, GSVAELP, introduces a hybrid GraphSAGE-VAE model that lever- ages local neighborhood aggregation with probabilistic latent-space embedding, suc- cessfully capturing both structural dependencies and latent relational patterns. This laid the foundation for robust structure-and-attribute-aware link prediction. The second study, MetaLP-DGI, introduced centrality-aware Deep Graph Infomax with meta-learning, enhancing embedding quality by incorporating influential node characteristics while improving generalization across heterogeneous networks. The third model, Hybrid Graph Embedding and Ensemble Learning, demonstrated that combining multiple embeddings with ensemble classifiers significantly improves predictive consistency and reduces model bias. vi Further, the fourth model enhancement is achieved through MetaLP-DGI, which utilizes Deep Graph Infomax (DGI) embeddings integrated with a centrality-aware transition matrix to capture both global and local structural dependencies. The meta- learning component in MetaLP-DGI optimizes the learning process across heteroge- neous datasets, improving robustness and adaptability. Complementing these in-depth approaches. The fifth study, Link Prediction in Social Networks: A Hybrid Approach with Graph Embedding and Ensemble Learning, combines structure- and attribute- based embeddings with ensemble classifiers, such as CatBoost and Random Forest, to deliver high-accuracy predictions in social network scenarios. Finally, the last study, UnifiedAttri2Vec–LSTM constructs a unified embedding by integrating multiple em- bedding algorithms through Attri2Vec and leverages LSTM to model temporal and structural dependencies simultaneously. Overall, this thesis contributes a comprehensive exploration of hybrid, generative, and meta-learning-based frameworks for link prediction, establishing a strong founda- tion for adaptive and scalable graph analytics. The progressive integration of centrality, attention, temporal evolution, and ensemble learning provides a unified roadmap for advancing intelligent link prediction in complex and dynamic networked systems.

2026-02-01T00:00:00Z CHAWLA, SONALI MALHOTR, RUCHIKA (SUPERVISOR) SHARMA, ANJALI (CO-SUPERVISOR) http://dspace.dtu.ac.in:8080/jspui/handle/repository/22754 2026-06-08T05:44:33Z 2026-03-01T00:00:00Z

Title: DEVELOPMENT AND VALIDATION OF HYBRID ALGORITHMS FOR SOFTWARE DEFECT PREDICTION Authors: CHAWLA, SONALI; MALHOTR, RUCHIKA (SUPERVISOR); SHARMA, ANJALI (CO-SUPERVISOR) Abstract: Software defect prediction (SDP) is an important research subject aimed at improving the reliability, maintainability, and overall quality of software systems. The rapid development of software projects raises the need for robust and accurate predictive models. While traditional machine learning (ML) and statistical methods have shown promise for SDP, challenges like high-dimensional data, imbalanced data, inefficient feature selection, and model-tuning limitations persist. To overcome these limita- tions, this research focuses on the development and validation of hybrid algorithms that leverage the power of both machine learning and metaheuristic optimization techniques to improve predictive performance capabilities for SDP. The research is validated through systematic review, empirical studies, and the development of novel algorithms applicable in real-world software development environments. The research is systematically structured into phases, addressing distinct compo- nents of SDP. The initial phase involves a synthesis of a systematic literature review that seeks to evaluate the latest hybrid algorithms that enhance the predictive perfor- mance of SDP models and identify research gaps. The review develops a framework for analyzing the current state-of-the-art with respect to hybrid algorithms on multiple dimensions and highlights the gaps that this thesis will work to address. In subse- quent phases, the research develops and validates several novel hybrid algorithms using benchmark datasets from repositories such as NASA, PROMISE, and AEEEM. These later phases include addressing the prime issues of dataset imbalance, design- ing improved feature selection techniques, implementing hyper-parameter tuning, and evaluating the proposed hybrid models against established baseline methods to demonstrate their effectiveness in real-world software defect prediction scenarios. The high-dimensional software datasets greatly influence the efficiency and ac- curacy of predictive models. Feature selection plays a vital role in simplifying complex datasets while retaining the most significant information. A hybrid SDP model integrating Binary Particle Swarm Optimization (BPSO), Synthetic Minor- ity Oversampling Technique (SMOTE), and Artificial Neural Network (ANN) is proposed to improve software quality. One of the significant contributions of this research is the development of a hybrid defect prediction framework that integrates filter feature selection(Information Gain, Relief F, and Chi-square) and metaheuristic optimization(Opposition-based Whale Optimization Algorithm) for feature selec- tion with attention-based deep learning classifier- Convolutional Neural Networks (1Dimensional- CNN), to achieve higher classification performance. This model is particularly valuable when dealing with large datasets, complex feature interactions, and the need for balancing multiple objectives, such as maximizing classification performance while minimizing the number of features. Predictive models for SDP often underperform when using default configurations, highlighting the critical need for hyperparameter optimization in maximizing model effectiveness. In this research work, we employed advanced optimization techniques, specifically Grey Wolf Optimization (GWO) and Salp Swarm Optimization(SSO) algo- rithms, in combination with machine learning and ensemble classifiers to create more effective hybrid models for SDP. These nature-inspired techniques navigate complex parameter spaces to achieve an effective balance between exploration and exploitation in an optimization process. This study highlights that appropriate hyperparameter tun- ing can yield a significant performance improvement because each predictive model undergoes comprehensive testing for different combinations of parameters before the optimal parameters are reached for each predictive model. Based on the promising outcomes of the hybrid algorithms developed for defect prediction, we further investigate their effectiveness by evaluating various hybrid approaches across multiple datasets to ensure the model 's generalizability. The experimental results are favourable for the hybrid models, which outperform traditional ML and statistical defect prediction models. This superiority is evident across key performance metrics, like F1-score, AUC-ROC, Recall, Precision, G-mean, and MCC. Furthermore, rigorous statistical testing confirms the reliability and robustness of these advanced techniques, reinforcing their effectiveness in SDP. In conclusion, this research significantly progresses the field of SDP by address- ing key predictive modelling challenges through the development and validation of sophisticated hybrid techniques. The study strengthens the effectiveness, reliability, and real-world applicability of defect prediction models. This study offers innovative methods for enhancing software quality, which benefits both academia and industry. The insights generated from this research provide a foundation for future advance- ments in predictive modelling, which will eventually help create software systems that are more dependable, efficient, and free of flaws.

2026-03-01T00:00:00Z