Using deep learning to generate suboptimal solutions for the Traveling Salesman Problem (TSP) while incorporating some randomness to improve the security of the tour path.

My project involves addressing a Multi-Agent Path Finding (MAPF) problem. The standard MAPF task entails determining paths for multiple agents in a given graph, leading from their current vertices to their respective targets, all while avoiding collisions with other agents. Simultaneously, the objective is to optimize a cost function associated with the paths.

Data privacy, using anonymization techniques such as k-anonymity or differential privacy to protect students’ identities from being disclosed. The key is to consider and balance the privacy utility trade-off.

Prove and prevent model collapse due to recource recommendation systems. Algorithmic recourse creates a feedback loop where users' strategic attempts to meet model criteria cause an endogenous data shift, making retrained models raise decision thresholds and inadvertently increase the difficulty of achieving successful recourse. We use continual learning and a data labeling algorithm to mitigate this effect.

How to generate images using a generative model (VAE), and find a method to explain the sources of these generated images through latent variables in the training data.

Analyzing the internal mechanisms of models, such as Denoising Diffusion Probabilistic Models, to investigate the variations and underlying regularities in neuron activations across diverse input distributions.

Proposing a geometric Explainable AI (XAI) framework that constructs enclosing simplices around query instances. By comparing a prediction against the consensus of its simplex vertices, we provide a systematic approach to audit individual fairness and uncover logical flaws in black-box models.