Applied AI systems need evaluation, context, and human consequences.

Adaptive Music Visualization and Offline RL

• offline RL
• affective computing
• FER
• EEG
• PPG
• human-centered evaluation

ProblemAdaptive creative systems need to respond to affective context without treating personalization as a purely aesthetic task.

Contribution boundaryQixuan drafted and revised manuscript sections on adaptive music visualization, synthesized literature across art therapy, cross-modal mapping, affective computing, and offline reinforcement learning, and helped frame a multimodal pipeline using FER, EEG, and PPG signals with conservative offline RL methods such as BCQ, IQL, and CQL.

TakeawayThis work shaped his view of AI as a system design problem: what data informs behavior, what adaptation is safe, and how a user-facing system should be evaluated.

AI-driven Therapeutic and Adaptive Media

• physiological signals
• personalization
• feature analysis
• adaptive systems

ProblemPersonalized therapeutic media depends on the connection between physiological signals, content features, and human response.

Contribution boundaryAs a research assistant at the University of Nottingham Ningbo China, Qixuan researched AI-driven art therapy and adaptive therapeutic media, extracted and analyzed music, visual, and physiological features, and supported refinement of machine learning pipelines for adaptive content generation.

TakeawayThe experience connects machine learning implementation with careful domain framing, especially where physiological data and personalization are involved.

Healthcare Prediction and Model Validation

• ROC/AUC
• calibration
• bootstrap validation
• LASSO
• model validation

ProblemPrediction models in healthcare can look strong on simple metrics while still failing on optimism, calibration, or generalization.

Contribution boundaryThrough directed reading and coursework, Qixuan studied overfitting, bootstrap validation, discrimination, ROC/AUC, calibration, shrinkage, LASSO, and model validation.

TakeawayThis background reinforces a practical evaluation habit: model performance has to be interpreted through validation design, not just headline accuracy.

LLM Privacy and Security Interests

• contextual integrity
• LLM privacy
• privacy attacks
• agent safety

ProblemLLM agents can expose sensitive context through training data extraction, personal data leakage, and hidden-task or prompt-driven behavior.

Contribution boundaryQixuan is interested in contextual integrity, privacy attacks and defenses, training data extraction, personal data leakage in LLM agents, and hidden-task chatbot privacy risks.

TakeawayThis interest gives his AI work a security and trust boundary: useful systems should also respect context, consent, and privacy.