Yinan Bu

Yinan Bu

Senior in Statistics (School of the Gifted Young, USTC). Summer research @ UMich Stats.

Research

Publication

Efficient synthetic network generation via latent embedding reconstruction.
F. Jiang, Y. Bu, S. Wu, G. Xu, J. Zhu.
(Under review)

Research Projects

Efficient Synthetic Network Generation via Latent Embedding Reconstruction
Jul. 2025 – Present
Advisors: Prof. Ji Zhu (Susan A. Murphy Collegiate Professor, Statistics, UMich) and Prof. Gongjun Xu (Professor, Statistics, UMich)

  • Developed a general, efficient framework for generating synthetic networks by combining latent space network models with a distribution-free generator over learned latent embeddings.
  • Built scalable pipelines for a diffusion-based latent embedding generator and a bootstrap-based latent embedding resampler, preserving key network characteristics while enabling efficient training with lower computational cost than many existing deep architectures(GitHub repository).
  • Conducted empirical studies on both simulated datasets and real-world datasets, showing that the proposed method efficiently generates networks that more faithfully preserve key characteristics than existing approaches.

Machine Learning and Hyperdimensional Computing
Apr. 2024 – Present
Advisor: Prof. Xueqin Wang (Chair Professor, Statistics and Finance, USTC)

  • Derived asymptotic information loss in vanilla Hyperdimensional Computing(HDC) and developed Hoeffding bounds for hypervector similarity and predictive accuracy.
  • Designed Feature-Subspace based Hyperdimensional Computing(FSHDC), a scalable model for fast classification and interpretation; applied to UK Biobank fMRI/MRI with +0.20 AUROC over vanilla HDC.
  • Integrated an attention mechanism into HDC training, improving accuracy by 30% on HAR vs. vanilla HDC and 15% vs. attention-only baseline.

Large Scale Optimization and GPU Acceleration
Jan. 2024 – Feb. 2025
Advisor: Prof. Xueqin Wang (Chair Professor, Statistics and Finance, USTC)

  • Worked on graph trend filtering (ℓ₁ minimization on graph differences) using ADMM; studied convergence vs. subproblem solvability trade-off.
  • Proposed Differential Operator Grouping–based ADMM(Doge-ADMM) with closed-form subproblems and parallel updates.
  • Built parallel implementations for first/second-order cases, achieving up to 30× speedup over existing methods (GitHub repo).

Academic Projects

Analysis of the Government Pension Fund of Norway (NBIM)
Jan. 2024 – Feb. 2025
Supervisor: Prof. Canhong Wen (Statistics and Finance, USTC)

  • Designed, implemented, and deployed an RShiny dashboard for the Norwegian Government Pension Fund Global (NBIM) with interactive Plotly charts and Leaflet world maps (Live demo, GitHub repo).
  • Conducted comprehensive analysis combining summaries, figures, and maps.

Uncertainty-Aware Time-Series Forecasting via Conformal Prediction
Dec. 2024 – Jan. 2025
Supervisor: Prof. Yu Chen (Statistics and Finance, USTC)

  • Reproduced the conformal prediction framework for probabilistic forecasting, building end-to-end calibration and evaluation pipelines.
  • Conducted experiments on AR/ARIMA, sales, air quality, and COVID-19 datasets, demonstrating robust uncertainty quantification with competitive interval widths and accuracy.