Zhimu Yang

Zhimu Yang

Theoretical Computer Science

Open to research opportunities

Research Interests
  • Algorithms & Complexity
  • Quantum Computing
  • Machine Learning
Programme III · Structured Learning

Graph Learning & Interpretability

A programme for explanations that respect graph structure, message-passing dynamics, and attribution principles.

☙ ❖ ❧
Central Thesis

A faithful graph explanation should reveal how information moves through the model, which structures sustain that movement, and why their contributions satisfy meaningful axioms.

Research Questions

When does an explanation reflect the model's actual message flow rather than a plausible subgraph found after the fact?

How should node, edge, feature, and substructure contributions be reconciled within one attribution framework?

Can architecture design remove the approximation error and evaluation cost of path-based attribution?

Research Record

FSX: Message Flow Sensitivity Enhanced Structural Explainer for Graph Neural Networks
Message-flow sensitivity · Cooperative games · arXiv:2601.14730
Paper
A Polynomial Architecture-Attribution Co-Design Framework for Exact Aumann-Shapley Attribution in GNNs
APEX · Exact attribution · arXiv:2607.21094
Paper Code
Graph Neural Network Interpretability

Research Map

Message-flow faithful explanation

Use internal information pathways to constrain the external substructures considered by an explainer.

Structural cooperative games

Develop contribution rules that account for interactions between graph components instead of treating them as independent features.

Architecture-attribution co-design

Build GNNs whose mathematical form makes complete and exact attribution computationally accessible.