Related to a project on explaining oversmoothing in Graph Neural Networks, we are experimenting on architectures and modifications that mitigate oversmoothing.
 
Details will be updated once the project reaches its final stages.

  State Space Models (SSMs), such as S4 and Mamba, have demonstrated exceptional performance in sequence modeling, with recent extensions to Graph Neural Networks (GNNs). A variant of SSM is the Liquid State Space Model (Liquid SSM), in which the update captures interactions across different parts of the input. Building on this liquid update equation and Mamba state update equation, we propose applying these and modify them to enhance representation building in GNNs.
 
Details will be updated once the project reaches its final stages.

  Transformers have revolutionized NLP with their self-attention mechanisms, but their quadratic complexity limits effectiveness for long-term context tasks. Mamba models, with linear complexity, achieve similar or superior performance while reducing computational overhead. This project aims to merge Mamba's efficiency with transformers' robust contextual understanding, optimizing them for long-context tasks.
 
Details will be updated once the project reaches its final stages.

  In this project, an oversampling technique was designed to tackle class imbalance within heterogeneous graphs. To tackle the issues, we will use components based on node-types within the oversampling technique.
 
Details will be updated once the publication reaches its final stages.

  Large-scale deep learning models are known for their large amount of parameters, weighing down on the computational resources. The core of the Lottery Ticket Hypothesis showed us the potential of pruning to reduce such parameters without a significant drop in accuracy. Quaternion neural networks achieve comparable accuracy to equivalent real-valued networks on multi-dimensional prediction tasks.
  Our experiments show that the pruned quaternion implementations perform better at higher sparsity than the corresponding real-valued counterpart, even in some larger neural networks.