我关注推荐系统、图神经网络和 RAG 应用开发，主要使用 Python 生态构建机器学习实验、后端 API 和可交互原型。

I focus on recommender systems, graph neural networks, and RAG-based applications. Most of my work is built with Python, PyTorch, PyTorch Geometric, FastAPI, and Streamlit.

• Graph neural networks for link prediction and candidate relation prediction
• Recommendation systems, recall/ranking pipelines, and explainable recommendation
• Retrieval-augmented generation for product search and decision support
• FastAPI / Streamlit applications for ML and AI prototypes

An improved SEAL-style graph neural network framework for link prediction. The project introduces an importance-aware subgraph sampling strategy based on ScoreGNN, reducing noisy neighborhood information during subgraph extraction.

Highlights:

• Reformulates node-pair matching as graph link prediction
• Uses DRNL structural labeling and DGCNN-style subgraph classification
• Adds ScoreGNN-based adaptive neighbor scoring and sampling
• Evaluates on Cora, CiteSeer, PubMed, GitHub, USAir, and C.elegans

Tech stack: Python, PyTorch, PyTorch Geometric, Jupyter Notebook, scikit-learn

An explainable e-commerce shopping assistant for musical instruments. It combines multi-channel retrieval, ranking, structured filtering, and review-based RAG to generate product recommendations with evidence.

Highlights:

• Natural language shopping intent parsing
• BM25, dense retrieval, LightGCN, and popularity recall
• Weighted RRF fusion ranking and hard filters
• Review evidence retrieval and recommendation explanation
• FastAPI backend and Streamlit frontend

Tech stack: Python, FastAPI, Streamlit, PyTorch, Qdrant, LangChain, OpenAI API

I am currently improving my portfolio around two lines of work:

1. Research-oriented graph learning projects, especially link prediction and recommendation recall.
2. End-to-end AI applications that combine retrieval, ranking, RAG, and explainable user-facing interfaces.

• GitHub: github.com/aqwddda
• Email: dzy.ivan.work@outlook.com