Projects
This space shows a few fun projects I built at SUSTech — spanning data science, machine learning, image processing, and more.
Bitcoin OTC Signed Trust Network: How a Reputation Market Forms Structure
Analyzing Bitcoin OTC as a signed temporal network, trust is reciprocal and clustered while distrust is sparse and cross-community; a small reputation core emerges, and a dynamic model predicts future ratings accurately and simulates network evolution with <4% error.
Time: 2026.05
Files: PDF Report
Meteorite Image Recognition via DINOv2-LoRA with Confidence Triage and Multimodal Fusion
Meteorite recognition using DINOv2‑LoRA with confidence triage and GPT‑4o‑mini review. It achieves public F1 = 0.84571 by routing only ambiguous samples to the VLM while 68% are decided by LoRA alone, effectively handling open‑world images under limited training data and domain shift.
Time: 2026.04-2026.05
Files: PDF Report
Data Agent System & AI Startup Blueprint
This project builds a data agent with reward‑based selection, Qwen3.5 fine‑tuning, and a Gradio demo, plus a lean startup plan for an AI resume generator featuring freemium model and high‑concurrency architecture.
Time: 2026.05
Files: PPT Presentation
OpenClaw for Clinical Data Analysis & Survival Analysis
This project employs OpenClaw as an autonomous agent for clinical data analysis, revealing key diabetes risk factors (BMI, HighBP, HighChol) and their interaction effects via Spark‑based ETL. It also reimplements survival analysis pipelines—Kaplan‑Meier, Cox proportional hazards, accelerated failure time, and customer lifetime value—on Spark, with SQL error case studies and a personal blog for publication.
Time: 2026.04
Files: PPT Presentation
Blog: https://zyhou.online
Competition Formats Diagnostics and Optimization for Dancing with the Stars
We infer hidden fan votes (93.1% accuracy), diagnose format biases (percent‑based amplifies fans, rank‑based stabilizes), quantify contestant/partner effects, and propose a robust LNZF+Judge Save system for fairer outcomes.
Time: 2026.02
Files: PDF Report
EduPal: LLM-Based Multi-Agent System for Educational Purpose
This project builds a complete educational LLM post-training pipeline: synthesizing SFT and preference data via structured instruction generation with quality/diversity filtering, fine‑tuning with LoRA and DPO, and deploying a collaborative multi‑agent teaching system (professor, TA, student, evaluator) using AutoGen to simulate university classroom interactions.
Time: 2025.11
Files: PPT Presentation
Distributed High-Frequency Tick Factor Computation on MapReduce
A MapReduce pipeline computes 20 quantitative factors from Shenzhen Level‑10 snapshots, using small‑file merging, in‑mapper state caching, and Combiner aggregation to achieve sub‑1% precision error and improved runtime efficiency.
Time: 2025.11
Files: PDF Report
RL Algorithms on CartPole‑v1
A comparative study of REINFORCE, A2C, and value‑based methods (NoisyD3QN, CQL) on CartPole‑v1, analyzing performance and stability. All achieve perfect scores with proper tuning; CQL shows robust offline learning. Highlights practical trade‑offs in sample efficiency and implementation complexity.
Time: 2025.11
Files: PDF Report
A Comprehensive Analysis of House Price Prediction
This study applies systematic preprocessing, feature engineering, and statistical testing to predict house prices. An ensemble of Lasso and XGBoost achieves the best CV RMSE of 0.115, identifying location, overall quality, and living area as key drivers.
Time: 2025.10-2025.11
Files: PDF Report
Credit Card Fraud Detection with Imbalanced Data
This study addresses severe class imbalance in credit card transactions using random undersampling and SMOTE oversampling. Logistic regression emerges as the best classifier among four models. A simple neural network is also implemented; undersampling yields better fraud recall while oversampling improves non‑fraud classification, highlighting the precision‑recall trade‑off.
Time: 2025.06
Files: PDF Report
Intelligent Scissors Image Segmentation Tool
This project implements an interactive image segmentation tool based on intelligent scissors. It applies gradient-based edge detection with Dijkstra/A* pathfinding, plus cursor snapping and path cooling for efficiency. The tool achieves extraction quality comparable to Photoshop's magnetic lasso, with user tests reporting high usability and significant time savings.
Time: 2025.05-2025.06
Files: PDF Report
Smartphone Price, Feature, and Rating Analysis
This project analyzes smartphone attributes—price, screen, camera, memory, and processor—to uncover relationships with brand and rating. Missing ratings are predicted via random forest, and performance radar charts highlight top models, with OnePlus and Huawei showing high excellence rates.
Time: 2024.12
Files: PDF Report