https://irju.jdvu.ac.in/jspui/handle/123456789/77 Sat, 07 Nov 2026 17:56:46 GMT 2026-11-07T17:56:46Z https://irju.jdvu.ac.in/jspui/handle/123456789/10823 Title: Mitigating class imbalance in indoor localization through autoencoder driven data augmentation Authors: Gayen, Bhaskar Abstract: Indoor Localization Systems (ILS) are crucial for applications such as asset tracking, emergency response, and indoor navigation. Despite significant advancements in the field, challenges related to accuracy and class imbalance in localization datasets remain. This thesis explores the application of generative models, specifically autoencoders, to address these challenges. By focusing on the JUIndoorLoc and UJIIndoorLoc datasets, we aim to create synthetic data that can balance the datasets and improve localization accuracy. The methodology involves tuning the autoencoder models according to the specific characteristics of each dataset. For JUIndoorLoc, collected from the 4th floor of a university building, the autoencoder is trained on data with 51 labels and 172 access points (APs). The UJIIndoorLoc dataset, from the IPIN2015 competition, presents a more complex scenario with 520 APs. Here, the autoencoder is trained on selected labels with specific sample instances to generate balanced synthetic data. Our results demonstrate that using autoencoders for data augmentation leads to significant improvements in localization accuracy. The training and validation loss metrics show that the autoencoder models effectively learn from the training data and generalize well to new data. This is crucial for enhancing the performance of ILS in real-world scenarios where data imbalance and limited labelled instances are common. Mon, 01 Jan 2024 00:00:00 GMT https://irju.jdvu.ac.in/jspui/handle/123456789/10823 2024-01-01T00:00:00Z https://irju.jdvu.ac.in/jspui/handle/123456789/10822 Title: Protein-protein interactions prediction, drug repurposing and molecular docking for emerging viral threats Authors: Paul, Debarati Abstract: PPIs are fundamental to the functioning of biological systems, regulating a myriad of cellular processes. Therefore, accurate prediction of PPIs is crucial for disease diagnosis and the development of novel or repurposed therapeutic interventions. Computational techniques streamline the forecast of PPIs and prove to be more cost-effective in terms of resources and time than experimental approaches. In this thesis, initially, a CNN-based novel methodology EnMAS has been proposed to predict PPIs by transforming amino acid sequences of interacting proteins into 2D images using a colour encoding scheme derived from evolutionary data such as PAM and BLOSUM matrices. Additionally, EnMAS has been applied to human- SARS-COV-2 PPIN to predict interactions between human hosts and the COVID- 19 virus. This is followed by another GNN-based study GSPPI where a topological feature of PPIN called graphlet (up to 5 nodes) has been leveraged to predict PPIs. Subsequently, in the next study, this graphlet feature and various amino acid composition-based features have been integrated to predict PPIs. This ensemble characteristic has demonstrated excellent prediction performance for PPI on the HIPPIE dataset. Owing to this performance, an illustrative case study has been conducted to predict the PHIs on the interacting protein network of MPXV and Humans. Mpox, a zoonotic illness triggered by MPXV presents a substantial risk to public health since it may be transmitted from person to person and there are no particular treatments available. Investigating how viral proteins interact with host proteins provides valuable insights into the dynamics of PHIs which is essential for designing effective vaccines. Utilizing the human protein targets for MPXV identified accurately, a drug repurposing study is undertaken to discover potential candidate drugs capable of curing Mpox. Additionally, validation has been conducted through subsequent docking studies on these promising drugs for MPXV using the prominent MPXV crystal structures currently available. The results of our study emphasize the effectiveness of using EnMAS and ensemble feature-based PPI prediction to understand the molecular processes involved in viral infection and to aid in the development of drugs for emerging viral threats such as Mpox and COVID-19. Though we have shown the efficacy of our method on Mpox as a case study, the approach is generalizable to a broad range of new emerging viral diseases with potential epidemic or pandemic threats, where the proposed method can reduce the risks and cost of developing new compounds and can offer a viable answer to the time constraints of traditional drug development. Mon, 01 Jan 2024 00:00:00 GMT https://irju.jdvu.ac.in/jspui/handle/123456789/10822 2024-01-01T00:00:00Z https://irju.jdvu.ac.in/jspui/handle/123456789/10821 Title: Prediction of landslide triggers in india using stacked ensemble models Authors: Das, Tirtha Abstract: Landslides are among the most devastating and recurrent natural disasters, impacting numerous mountainous regions worldwide. India is no exception, with frequent landslides affecting vital economic sectors such as transportation and agriculture, often resulting in significant loss of life. A landslide trigger plays a crucial role in initiating the downward movement of rock, earth or debris on a slope. Triggers include natural events like heavy rainfall, earthquakes and rapid snowmelt, as well as human activities like construction, deforestation and mining, which disrupt slope stability and cause significant damage and loss of life. Accurately identifying the triggers of landslides is essential for reducing risks and mitigating their impact. Many researchers have explored landslide prediction, but only a few have focused on landslide triggers. The main challenge with India’s landslide dataset is its inherent imbalance. To address this issue, the Synthetic Minority Over-sampling Technique (SMOTE) was employed. SMOTE generates synthetic examples for the minority classes, thereby balancing the dataset. This approach improves the model’s ability to correctly identify and predict landslides triggered by less frequent events, leading to more robust and reliable predictions across all classes. To predict landslide triggers, a stacked-based machine learning model has been developed, which outperformed other baseline models with an accuracy of 94.6% on unseen data. This achievement underscores the effectiveness of the stacked ensemble model in capturing the complex relationships between various factors contributing to landslide triggers. The robust performance of the stacked ensemble model highlights its potential as a valuable tool for improving landslide prediction and enhancing public safety. Mon, 01 Jan 2024 00:00:00 GMT https://irju.jdvu.ac.in/jspui/handle/123456789/10821 2024-01-01T00:00:00Z https://irju.jdvu.ac.in/jspui/handle/123456789/10820 Title: Sentimental symphony: maximizing gpt capabilities for text generation and emotional insight Authors: Maity, Soumik Abstract: In recent years, Natural Language Processing (NLP) models, particularly Generative Pre-trained Transformers (GPT), have demonstrated remarkable capabilities in generating coherent and contextually relevant text. However, while these models excel in linguistic proficiency, they often fall short in capturing the nuanced emotional undertones present in human communication. This paper introduces "ECGText" (Emotionally Controlled Generation of Text) a novel model designed to enhance text generation by integrating emotional intelligence into GPT-based systems. By leveraging advances in sentiment analysis and emotion recognition, Sentimental Symphony empowers GPT to imbue generated text with appropriate emotional nuances, thereby fostering more engaging and relatable communication. We present a comprehensive overview of the framework, discussing its architecture, training methodology, and evaluation metrics. Additionally, we conduct experiments on diverse text generation tasks, showcasing the efficacy of Sentimental Symphony in producing emotionally resonant outputs across various domains. Our results demonstrate significant improvements in generating text with nuanced sentiment, paving the way for more emotionally intelligent AI systems that better mirror human expression and understanding. Mon, 01 Jan 2024 00:00:00 GMT https://irju.jdvu.ac.in/jspui/handle/123456789/10820 2024-01-01T00:00:00Z