Zhang Jinshuai

Zhang Jinshuai

Harbin Institute of Technology

Research and Project Experience

中文版本

Anomaly Score-Guided Feedback in Graph Anomaly Detection (FGAD) 2024.10 - 2025.03

  • Affiliation: INET Laboratory, Harbin Institute of Technology (Weihai)
  • Journal Submission: First-author pre-submission to CCF-B conference ICDM 2025.
  • Anomaly Feedback Mechanism: Improved traditional GCN embedding layers by constructing a lightweight GCN architecture with feature diffusion weights. Adjusted node feature aggregation based on anomaly scores, structural features, and attribute similarity to suppress noise propagation and enhance model robustness.
  • Graph Anomaly Detection: Implemented a GAE-based detection model combining node reconstruction (attribute anomalies) and structural reconstruction (topological anomalies).
  • Data Partition Optimization: Enhanced mini-batch partitioning using Metis and k-means clustering to reduce edge information loss.
  • Performance: FGAD model outperformed 11 baseline methods on 7 anomaly detection datasets (Cora, Flickr, Reddit, Books, etc.).

Scalable Generative Tool Calling via Structure-Aware Semantic Tokenization 2024.11 - 2025.05

  • Affiliation: NLPR Laboratory, Chinese Academy of Sciences
  • Journal Submission: EMNLP 2025 Findings (Second-author).
  • Theoretical Research: Explored incremental learning (DSI++), collision-free codebook methods, and IDGenRec alternating training strategies.
  • Data Processing: Addressed Tool Agent data scarcity by constructing a 589k multimodal Tool dataset covering 46k tool categories to optimize Toolkengpt’s tool comprehension.
  • Model Training: Reproduced and optimized the Toolkengpt model, conducting experiments on generated datasets to evaluate NDCG and accuracy.
  • Model Fine-tuning: Participated in DeepSeek-R1 fine-tuning research, building a 1.77M dataset based on MoleculeNet and achieving 98.95% classification accuracy.

Microservice Fault Root Cause Localization 2024.07 - 2025.03

  • Affiliation: ICES Laboratory, Harbin Institute of Technology
  • Methodology Study: Reproduced graph anomaly detection (e.g., DONE) and microservice fault tracing (e.g., APG) papers, defining problems and collecting data.
  • Data Processing: Processed microservice datasets and implemented data augmentation methods, constructing timestamp-based snapshot graphs as DGL Datasets.
  • Model Optimization: Refactored anomaly-aware graph embedding models by migrating frameworks from PyG to DGL and adopting lightning-hydra-template to improve training efficiency.

Early Warning System for Risk Propagation in Industrial Chains Based on GNNs 2024.10 - 2024.11

  • Role: Assistant developer for risk assessment model and path extraction algorithms.
  • Data Processing: Handled large-scale enterprise node and edge relationship data using one-hot encoding and timestamp-based dataset partitioning.
  • Model Design: Built a GAT-based heterogeneous graph neural network integrating node/edge features for enterprise risk probability prediction.
  • Path Extraction: Developed attention-based path tracing algorithms to visualize risk propagation paths.
  • Optimization: Enhanced model performance on complex graph structures by introducing multi-head attention and edge feature fusion.

A Blockchain anomaly transaction detection system using GNNs (BCWatch) 2024.03 - 2024.08

  • Role: Team Leader (17th National Information Security Competition)
  • Model Optimization: Reduced computational complexity through lightweight graph representation learning and clustering architectures, enabling large-scale graph processing under resource constraints.
  • Clustering Precision: Unified graph representation learning and clustering via node discrimination and expansion-contraction loss, achieving “intra-class minimization and inter-class maximization.”

LLM Practical Training Program 2024.04 - 2024.05

  • Learning Content: Studied LLM fundamentals, prompt engineering, and fine-tuning techniques.
  • Project: Developed the “Emotion Cube” emoji generator on ModelScope, obtaining the CSTP LLM Application Engineer Certification.

AI Research Project at Nanyang Technological University 2023.01 - 2023.03

  • Focus: AI applications in healthcare.
  • Outcome: Received completion certificate and recommendation letter from Prof. Teoh Teik Toe.