JZ 中文

Recommendation Algorithms · Machine Learning

Tsinghua University seal

Tsinghua University

M.E. in Electronic Information (085400)

Institute for Network Sciences and Cyberspace · 2026-Present

Jinshuai Zhang

I build ranking and learning systems for practical recommendation problems, supported by research in scalable tool calling and graph machine learning.

Email GitHub
Portrait of Jinshuai Zhang
Recommendation systems · Tool calling · Graph learning
Distinctions

01

Distinctions

Selected national awards and individual honors.

Competition Awards

  • National Grand Prize ICBC Cup National College Student Financial Technology Innovation Competition
  • National First Prize iCAN AI Challenge
  • National Third Prize · Team Lead National College Student Information Security Contest, Works Competition

Personal Honors

  • National Honor National Scholarship
  • University Honor 23rd Top Ten Student and Mazu Guang Scholarship
  • University Honor Model Outstanding Student Cadre
  • Graduate Honor Outstanding Graduate

02

Selected Work

Industry ranking work and research projects spanning tool agents, graph learning, and risk intelligence.

01 Industry Recommendation

ByteDance · Recommendation Algorithm Intern · 2026

Mini-game Recommendation Ranking

Framed a large-scale recommendation ranking problem around mini-game candidate supply and anchor conversion. Built privacy-safe samples, aligned labels and statistical definitions, and applied controlled data checks to make offline evaluation reliable before model iteration.

Developed a DNN ranking workflow spanning sample construction, candidate expansion, personalized reranking, ablation studies, and controlled online validation. The first version established a click-propensity ranking loop; later iterations evaluated deeper launch and stay-depth objectives through multi-task learning, long-tail regression, and warm-start strategies.

Owned the implementation path from offline analysis to online serving. The engineering work included feature consistency, retrieval filtering, layered observability, and reusable troubleshooting practices across data, retrieval, model, and serving stages.

02 Scalable Tool Calling

Findings of EMNLP 2025 · Second author

ToolScaler: Scalable Generative Tool Calling via Structure-Aware Semantic Tokenization

The team method represents roughly 47k API functions from ToolBench as composable, structure-aware semantic token sequences. Using semantic compression, related tools receive nearby codes, while constrained generation reduces selection cost across an exceptionally large tool space.

The framework combines initial code construction, iterative training, and post-guided refinement to learn tool semantics and generalize to tools unseen during training, rather than depending on a fixed retrieval inventory.

My contribution focused on processing ToolBench documentation, query-tool pairs, and invocation trajectories; constructing and quality-checking high-quality training data; supporting training; and running retrieval, unseen-tool generalization, and end-to-end calling experiments with result analysis.

03 Graph Anomaly Detection

First author · Manuscript in preparation, target: IEEE TDSC

FGAD: Feedback-Guided Anomalous Diffusion Suppression for Graph Anomaly Detection

Addressed the spread of anomalous information through graph neighborhoods and its contamination of normal-node representations. The method combines attribute reconstruction with edge-level structural consistency to produce feedback signals during message passing.

An MLP-based adaptive diffusion matrix and decoupled graph convolution regulate propagation, while Metis subgraph partitioning supports scalable mini-batch training. Ablation studies isolate the effects of feedback, suppression, and structural components.

As first author, led method design, implementation, experiment construction, and analysis. Against 11 baselines, FGAD achieved the best AUC on 7 of 8 evaluated datasets, including approximately 92% on Inj Flickr.

04 Financial Risk Intelligence

National competition project · Team collaboration

ZDC: Intelligent Anti-Money Laundering Detection Platform

Built a graph-based anti-money-laundering research prototype for a national competition. It jointly models entities and transactions, using attribute and structure reconstruction to identify anomalous relationship patterns.

The team approach combines anomaly feedback, contrastive learning, and attention-based enhancement, and produces interpretable evidence paths to support analysis rather than only isolated risk scores.

My role centered on model training and method innovation, including experiment tuning and interpretation-oriented algorithm integration. The competition research prototype received the ICBC Cup National Grand Prize.

05 Service Reliability Research

ICES Research Intern · Research and engineering prototype

Graph Anomaly Detection for Microservice Root-Cause Localization

Studied microservice root-cause localization by reproducing and comparing representative graph anomaly detection baselines, with parameter and evaluation settings organized across multiple datasets.

Reconstructed a DONE-style dual-branch training framework with DGL and PyTorch. Edge-aware aggregation combines node attributes with edge observations while preserving separate structural and attribute branches.

My work covered data preparation, baseline implementation, experiment comparison, and engineering integration. The result is a research and engineering prototype, not a deployed root-cause analysis system.

03

Experience

  1. ByteDance

    Recommendation Algorithm Intern · Mini-game ranking

    • Built privacy-safe samples and evaluated DNN ranking, candidate expansion, and controlled validation choices.
    • Extended the modeling path from click propensity toward launch and stay-depth objectives, while supporting online integration, observability, and cross-stage troubleshooting.

  2. Institute of Automation, Chinese Academy of Sciences

    NLPR Research Intern · Tool agents and unseen-tool generalization

    • Processed ToolBench documentation, query-tool pairs, and invocation trajectories into quality-controlled training data.
    • Supported training and analyzed retrieval, unseen-tool generalization, and end-to-end calling experiments.

  3. Harbin Institute of Technology

    ICES Research Intern · Graph anomaly detection for root-cause localization

    • Reproduced graph-anomaly baselines and organized multi-dataset experiment comparisons.
    • Reconstructed DONE-style dual-branch training in DGL and PyTorch with edge-aware aggregation.

04

Education

2026.09-Present

Tsinghua University

M.E. Student in Electronic Information (085400) · Institute for Network Sciences and Cyberspace

2022.09-2026.06

Harbin Institute of Technology (Weihai)

B.Eng. in Cyberspace Security · School of Computer Science and Technology · Rank 2/75 · GPA 94.09/100 · CET-6

  • Database Systems (99) · Computer Organization (99) · Computer Networks (98)
  • Data Structures and Algorithms (96) · Operating Systems (95)

05

Let’s build better ranking systems.

I am open to opportunities and conversations in recommendation algorithms, machine learning, and applied research.

zhangjinshuai0711@163.com github.com/zhangjinshuai0711