Xuan Zhang

• Ph.D. Computer Science student with research interests in Applied Machine Learning and Applied Deep Learning, specializing in Discrete Representation, Reinforcement Learning (RL), Graph Neural Networks (GNN), and Natural Language Processing (NLP) with applications to real-world scenarios including Finance and Computer Vision. Current research focuses on Portfolio Management, Time Series Forecasting, Fake Image Detection and synthetic data generation for rare events.
• Fast learner, self-motivated, and collaborative professional with the ability to think critically and deliver analytical solutions clearly, concisely, and in a timely manner. Combines 6+ years of industry software engineering experience with advanced research expertise

Extension of Adaptive and Explainable Margin Trading via LLMs and Reinforcement Learning (March 2025 - Present) Cross-Sectional Portfolio Optimization on Dynamic Network Structure (May 2025 - Present) Enhancing Vision Transformer Robustness through Minimum Perturbation Distance Regularization (January 2025 - August 2025)

• Integrating Large Language Models for enhanced market context understanding and sentiment analysis in trading decisions
• Implementing advanced Reinforcement Learning techniques for dynamic portfolio management and risk assessment
• Developing transparent decision-making frameworks for algorithmic trading adaptability and explainability
• Technologies: Large Language Models, Reinforcement Learning, Portfolio Optimization
• Developing dimensionality reduction techniques with economic interpretability for financial factor modeling
• Implementing dynamic factor integration methods for adaptive portfolio structures across market conditions
• Building graph-based models for cross-sectional portfolio analysis using network topology
• Technologies: Graph Neural Networks (GNN), Portfolio Optimization, Subgraph Analysis, Factor Models
• Theoretical Framework Development: Established rigorous mathematical framework for analyzing minimum perturbation distance (d*) in VQ tokenization, deriving explicit formulas based on Voronoi cell geometry and connecting to adversarial robustness guarantees
• Novel Regularization Method: Developed differentiable d* regularization technique applicable to any vector quantization training framework (VQ-VAE, ViT-VQGAN, VQGAN), with efficient approximation strategies for computational tractability
• Comprehensive Experimental Validation: Conducted extensive experiments on ImageNet100 and ImageNet-C-100 datasets, demonstrating significant robustness improvements while maintaining reconstruction quality and downstream task performance
• Codebook Geometry Analysis: Analyzed how d* regularization affects codebook structure, token utilization patterns, and geometric properties, providing insights for future discrete representation learning research
• Technologies: Vision Transformers, Vector Quantization, Adversarial Robustness

Peer-reviewed journal and conference papers (* equal contribution)

Under Review

• Jingyi Gu* , Xuan Zhang* , Guiling Wang. "Beyond the Norm: A Survey of Synthetic Data Generation for Rare Events." ACM Computing Surveys. https://arxiv.org/pdf/2506.06380
• Xuan Zhang, Jianyi Yang, Junyi Ye, Martin Renqiang Min, Guiling Wang, Zachary Izzo. "Enhancing Vision Transformer Robustness through Minimum Perturbation Distance Regularization." AAAI 2026,

Chair, ACM-Women NJIT Chapter -YWCC (2025-05-01 -  Present)

Leading initiatives to support and promote women in computing fields.

Organized the event "NJIT ACM-Women 10-Year Anniversary Celebration"

• AWS Certified Developer – Associate
• AWS Certified Cloud Practitioner