Thesis Proposals

MAster Degree in Computer science - Data science - CYBERSECURITY - Computational Finance

• Mining Users’ Attributes from their Public Spotify Data Using Graph Neural Networks

The proliferation of data on digital platforms like Spotify offers a unique opportunity to delve into the intricacies of user preferences and behaviours. Spotify's extensive interconnected data structures encompassing users, songs, playlists, artists, and more present a wealth of information. However, extracting latent user attributes from these intricate network structures poses a significant challenge. Traditional machine learning techniques struggle to unravel the complex relationships and dependencies inherent in such networks, underscoring the need for advanced methodologies such as Graph Neural Networks (GNNs). This thesis proposes harnessing the power of Graph Neural Networks (GNNs) to uncover sensitive user attributes from Spotify data. Specifically, our focus is on discerning recurring musical characteristics associated with users' individual attributes, including demographics, habits, or personality traits. 

For further details: https://arxiv.org/abs/2401.14296

In collaboration with: Mauro Conti (SPRITZ Research Group), Luca Pajola (Spritzmatter), Pier Paolo Tricomi (Spritzmatter)

• Hypernetworks for graph-structured data 

Graph-structured data is ubiquitous in various domains, including social networks, biological networks, and knowledge graphs. Graphs are dynamic, heterogeneous, and can vary significantly in size, making traditional neural network architectures less suitable for effective learning. Hypernetworks, a class of neural networks that generate weights for other networks, have shown promise in capturing complex patterns in high-dimensional data. Hypernetworks offer a unique approach by generating weights dynamically, potentially addressing these challenges. However, their application to graph-structured data is an emerging area of research that requires in-depth exploration. The goal of the thesis is to explore the application of Hypernetworks to graph-structured data, aiming to develop innovative techniques that enhance scalability, flexibility, and performance in graph learning tasks.

• Speech enhancement in multi-talker environments

In everyday environments, speech signals are often contaminated by various interfering noises and multiple speakers, making it challenging for speech recognition systems and communication devices to function effectively. Multi-talker environments, where multiple speakers are active simultaneously, pose a significant problem for speech enhancement techniques. Traditional speech enhancement methods struggle to effectively isolate target speech from interfering speakers and background noise in multi-talker environments. The thesis project aim to investigate advanced deep learning approaches to enhance speech signals in such complex acoustic scenarios. The research focuses on developing innovative methods for speech separation, denoising, and quality improvement in multi-talker environments.

• Graph Neural Networks for Large-Scale Datasets

Graph Neural Networks (GNNs) have emerged as powerful tools for processing graph-structured data, exhibiting promising performance in various applications, but handling large-scale graphs poses significant challenges due to their size and complexity. Existing GNN models often suffer from scalability issues, limiting their applicability to real-world, large-scale datasets. The thesis goal is to explore and address these challenges by proposing novel techniques to optimize GNNs for processing massive graphs efficiently while maintaining high prediction accuracy.

• Deep Learning Models for Neuroimaging Analysis 

Neuroimaging plays a crucial role in understanding the structure, function, and disorders of the brain. With the advent of Deep Learning techniques, there has been a paradigm shift in the field of neuroimaging analysis. Deep Learning models have demonstrated exceptional capabilities in extracting intricate patterns from neuroimaging data, leading to more accurate diagnoses and personalized treatment strategies. This thesis project aims to explore the application of Deep Learning models in neuroimaging analysis, focusing on developing models that can handle diverse imaging modalities, interpret complex brain structures, and provide reliable diagnostic insights. 

• Forecasting Stock Prices Using Deep Learning

Deep learning, particularly Recurrent Neural Networks (RNNs), has emerged as a potent tool for various predictive tasks, yet its application to stock price prediction presents unique challenges due to the intricate nature of financial markets. While existing deep learning methods exhibit promise, their effectiveness in capturing the nuanced patterns of stock price movements requires further exploration. This thesis aims to investigate the application of RNNs and other deep learning techniques in stock price prediction and explore novel approaches to enhance their predictive accuracy.