Howdy! I’m a Ph.D. candidate in Statistics at Boston University, co-advised by Prof. Debarghya Mukherjee and Prof. Luis Carvalho, and I also collaborate with Prof. Nabarun Deb. Before BU, I earned my M.A. in Statistics from Columbia University and my B.S. in Mathematics from Shandong University, including a year of joint training at the Academy of Mathematics and Systems Science(AMSS), Chinese Academy of Sciences. My research sits at the intersection of statistics and machine learning, where I develop theoretically grounded transfer-learning and representation-learning methods—spanning optimal transport, graph mining, multimodal learning for structured, heterogeneous data in low-sample, high-dimensional, and non-IID settings.

The question that keeps me up (in a good way):

How can we reuse past knowledge when the world—and the data—won’t sit still?

In statistical learning, this is about transferring geometry or smoothness from a well-understood source distribution to a smaller, noisier target under shift. In reinforcement learning, the source might be prior trajectories, simulators, or related tasks, while the target is the evolving environment, so we need principled rules for what to keep, what to adapt, and what to forget. And yes! LLMs/VLMs make this even more exciting (and tricky): they already contain a lot of cross-domain knowledge, but the real challenge is extracting and specializing it safely for downstream tasks without overfitting, hallucination, or misalignment.

What I build

Theory that supports practice Minimax rates · oracle inequalities · regret bounds · safe-transfer criteria under covariate or structural shift.	Graph-structured transfer Aligning and transporting information across graphs and manifolds — robust transfer when correspondence is messy or unknown.
RL & bandits under drift Warm-started policies with uncertainty-aware adaptation for reliable sequential decision-making in changing environments.	Transfer for LLMs / VLMs Controlled adaptation · domain grounding · structure-preserving fine-tuning — so models adapt without getting sloppy.

Curious about my research? I’ve put together beginner-friendly slide decks on my main research directions: transfer learning, graph learning, optimal transport, and LLMs for time series.

Along my academic journey, I have been deeply fortunate to study and conduct research under the guidance of inspiring scholars, including Prof. Zhanxing Zhu, whose influential work includes Spatio-Temporal Graph Convolutional Networks (STGCN) for traffic forecasting, and Prof. Yongshun Gong. Their perspectives on deep learning, representation learning, and structured spatio-temporal systems have profoundly shaped how I think about evolving, heterogeneous data, and have guided my pursuit of principled transfer learning methods.

Beyond theory and modeling, I am drawn to building AI applications that reflect how I see people and the world. I have always felt that human beings are more than their outward forms, that something of the spirit, memory, and inner life exceeds the body that temporarily carries it. That is why I am especially fascinated by cinema, atmosphere, and emotionally resonant digital experiences ✨

🔥 News

2025.09: 🎉 My first-author paper “Transfer Learning on Edge Connecting Probability Estimation Under Graphon Model” is accepted by (NeurIPS 2025)!
2025.08: 🎉 My co-authored paper “Cross-Domain Hyperspectral Image Classification via Mamba-CNN and Knowledge Distillation” is accepted by (IEEE TGRS 2025)!

📝 Publications

Leading Author

Transfer Learning on Edge Connecting Probability Estimation Under Graphon Model

First graphon-level transfer without node correspondence — aligns graphs via Gromov–Wasserstein and transfers edge structure nonparametrically.
Residual smoothing unlocks small/sparse targets with convergence & stability guarantees; SOTA on link prediction and graph classification.

Phase Transition in Nonparametric Minimax Rates for Covariate Shifts on Approximate Manifolds

New minimax theory for "near-manifold" shift: exposes a sharp phase transition controlled by the support gap between target and source neighborhoods — unifying multiple geometric-transfer regimes.
Ratio-free, adaptive estimator: achieves near-optimal, dimension-adaptive rates without density ratios and without assuming known geometry (works under approximate manifold mismatch).

From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space

Bridges the text–time-series modality gap: introduces a Temporal Evolution Semantic Space (TESS) that distills free-form text into interpretable temporal primitives (mean shift, volatility, shape, lag), instead of directly fusing noisy token embeddings.
LLM-guided yet numerically grounded forecasting: uses structured prompting + confidence-aware gating to inject reliable semantic signals as prefix tokens into a Transformer forecaster, yielding robust gains under event-driven non-stationarity (up to 29% error reduction).

SCOT: Multi-Source Cross-City Transfer with Optimal-Transport Soft-Correspondence Objectives

Sinkhorn entropic-OT coupling enables many-to-many region alignment across cities — no node matching required.
OT-weighted contrastive loss + target-aware prototype hub prevents collapse and scales cleanly to multi-source heterogeneity.

INCM: INConsistency-aware Multi-modal Recommendation with Cross-Modal Hard Negatives

Inconsistency-aware multimodal ranking: studies how cross-modal discrepancies may provide complementary ranking evidence or degrade fusion quality — explicitly modeled in training.
Cross-modal hard negatives + synergy-aware ranking loss: proposes CHNS to mine modality-specific hard negatives across branches, and a Synergy-aware BPR loss to ensure the fused branch achieves stronger preference margins than unimodal branches.

Co-author

Cross-Domain Hyperspectral Image Classification via Mamba-CNN and Knowledge Distillation

Hybrid spectral–spatial modeling for domain shift: integrates a Mamba-based global spectral encoder with CNN local feature extraction, capturing long-range dependencies while preserving fine-grained spatial structure.
Dual-level transfer via distillation + graph alignment: performs teacher–student knowledge distillation for distribution alignment and OT-guided graph consistency across domains, yielding robust cross-domain generalization under severe spectral mismatch.

Semantic Scientific Graph Pruning for Reliable Agentic Paper Reproduction

SSGP prunes dense scientific graphs into task-adaptive subgraphs via rank-based ensemble scoring — drastically shrinks agent search space.
Reuse–patch execution + confidence-weighted aggregation boosts reproducibility, stability, and success rate of LLM scientific agents.

🤖 LLM Engineering Projects

AlignDPO _{DPO · IPO · KTO · QLoRA · Mistral-7B · HH-RLHF}	RAGAudit: Hallucination Detection _{BM25+FAISS · NLI · SelfCheckGPT · sem. entropy · Mistral-7B}
Congestion Pricing Analyzer _{TWFE · CS-DiD · Synth DiD · Double ML · 12M+ NYC TLC}	CausalLens: LLM-Augmented Causal Pipeline _{DoWhy · Double ML · Causal Forest · Claude API · Streamlit}
GraphRAG: Multimodal RAG _{dense + entity graph + CLIP · FastAPI · ChromaDB}	Adaptive RAG _{query routing · iterative retrieval · self-check · FastAPI}
DraftVerify: Speculative Decoding _{draft + verifier · latency · throughput · acceptance}	HQQ: 1-bit Quantization _{1–8 bit · proximal opt · W1G64: 12.7× · >4× speedup}
Causal Promotion Optimization _{AIPW · LightGBM · DRLearner CATE · OR-Tools · FastAPI}	Demand Forecasting _{Seasonal Naive · LightGBM · TFT · M5 · 28-day · store-SKU}

📖 Educations

2021.09 – Now: Ph.D. in Statistics, Boston University
2019.09 – 2020.05: M.A. in Statistics (Data Science Track), Columbia University
2018.05 – 2019.06: B.S. in Mathematics, Chinese Academy of Sciences (Jointly Supervised Talent Program)
2015.09 – 2019.06: B.S. in Mathematics, Shandong University

💻 Internships

Plymouth Rock

Data Scientist Intern · Plymouth Rock Insurance
_{📍 Boston, MA · 🗓️ May 2025 – Aug 2025}

Architected an end-to-end AWS SageMaker pipeline for property-level loss prediction using an XGBoost Tweedie model on multi-million-policy data, lifting Gini by +4.3% over the production baseline and directly improving underwriting risk segmentation.
Pioneered an LLM-powered visual risk scoring system combining GPT-4o multimodal reasoning with Google Street View imagery to capture previously unobservable property features (roof condition, surroundings, hazards); integrated outputs into downstream actuarial pricing models as a novel signal layer.
📎 For a high-level, non-confidential summary of this work, see the Home Insurance slides.

✨ My Apps

A quiet collection of cinematic, atmospheric, and emotionally resonant side projects — part digital keepsakes, part memory-keepers. See all →

_{🌲 Wilderness}	_{✨ MBTI Vibe}	_{🎬 What If Cinema}	_{✉️ Letters from Screen}	_{✈️ If You Disappeared}
_{🎟️ Souvenirs}	_{🗺️ Map of Me}	_{🦋 A Room in Macondo}	_{✒️ Say It Like a Classic}	_{🏛️ Boston Archive}

🎖 Honors

2025: Student Travel Grant, Boston University
2025: Ralph B. D’Agostino Endowed Fellowship, Boston University
2025: Outstanding Teaching Fellow Award, Boston University
2019: Outstanding Graduate, Shandong University
2018: Hua Loo-Keng Scholarship, Chinese Academy of Sciences
2018: National Gold Award, Internet+ Innovation & Entrepreneurship Competition
2018: First-Class Scholarship, Shandong University
2018: Outstanding Student Leader, Shandong University

📂 DS Projects

Dog Classification _{VGG16 · ResNet50 · Flask · 75.48%}	Credit Risk _{XGBoost · SMOTE · AUC 0.976}	Pedestrian Detection _{Fast R-CNN · Siamese · few-shot}
Financial Sentiment _{DistilBERT · 85% · 30%↑ speed}	Mask Detection _{ResNet50 · Grad-CAM · 94%}	Spam Detection _{TF-IDF · NB · P 96 / R 94}
Airbnb Dashboard _{R Shiny · maps · filtering}	Bayesian Logistic _{RStan · Spike-and-Slab · MCMC}	A/B Testing _{Bootstrap · power · +15% conv.}
Time Series Forecast _{SARIMA · ETS · Prophet}	Movie Recommendation _{ALS · SVD · +15% / −20%}	Customer Segmentation _{K-Means · elbow · silhouette}

📝 Service & Teaching

Presentations · CIKM 2024, NeurIPS 2025
Reviewer · CIKM 2025, ICME 2026, ICML 2026, KDD 2026
Instructor @ Boston University · MA 582 Mathematical Statistics, MA 113 Elementary Statistics
TA @ Boston University · MA 575 Generalized Linear Models, MA 582, MA 415 Data Science in R, MA 214 Applied Stats

🎨 Interests

🎵 Mandarin R&B loyalist — Leehom Wang, David Tao, Khalil Fong🦋, Dean Ting

🎹 Trained in piano, calligraphy, and ink painting

🏞️ National park lover · 🫧 lake admirer · 🌅 opacarophile — welcome to my Gallery

Yuyao Wang