• Recruiting: I am always looking for motivated students, especially Ph.D. students starting at RIT in Spring 2027 or Fall 2027.
  • Before reaching out, please read the Work With Me page for details.

Research Interests

I received my Ph.D. from Virginia Tech in 2020, advised by Prof. Chang-Tien Lu. I previously served as an Assistant Professor at Mississippi State University before joining RIT. My work runs along three lines, in order of weight.

1Network science, graph machine learning and dynamical systems

How do behaviors spread, settle and break in connected systems, and how can we estimate and control them? I look for generalized patterns that hold across very different networks and for cross-system behaviors that pass from one network to another, in traffic grids, power networks, epidemics, social networks, banking and supply chains. My tools are spectral and functional analysis and uncertainty quantification (UQ).

Relevant support: NSF CAREER #2443266 · NSF CIRC #2345921 · all grants →

Generalized patterns across systems

① "Analogy learning": Generalized patterns across systems. Shared operators, conserved quantities, and structural invariants that transfer across unrelated graph domains.

Cross-system behaviors

② "Interplay learning": Cross-system behaviors. Behaviors that propagate between networks, where cross-edges carry the dynamics.

③ "Network Digital Twin": Executable replicas of coupled infrastructure. Traffic, power, and communication testbeds for stress-testing estimation and control.

2Mechanistic understanding of LLMs

A model’s activation graph before and after a knowledge edit, with changed and lost edges marked DELTA GRAPH before after edit
Delta-graph signatures (SDM 2026) read a model’s activations as a graph and flag knowledge edits that pass standard checks but damage unrelated facts. The talk From Networks to Neural Networks presents this same work.

3Computational musicology

1 Learning. The ball walks Beethoven's melody at his tempo. Each move draws an arrow, thicker when it repeats. 1/2 means half of this note's moves so far went there.

2 Composing. Each circle turns into a spinner. A wedge's size is how often that note led to the next one. The needle lands at random, so common moves win more often.

A first-order Markov chain graph counts which note follows which in the opening of Für Elise. Each node holds one note's transition probabilities, and spinning them writes a new tune.
A first-order Markov chain graph learned from Für Elise: each node spins its transition probabilities to write a new tune (click to listen). One graph per piece → Earlier work fused genres with adversarial dual learning (ICDM 2017, with Alexander Lerch).

Selected Paper

Full publication

Cross-Network Dynamics

Network Interdiction Goes Neural
Lei Zhang, Zhiqian Chen, Chang-Tien Lu, and Liang Zhao
Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2025 CORE A*

Uncertainty Quantification

Spectral Graph Theory

Spatial Forecasting

Software / Hardware

Symbolic Discovery

Science

Music

Selected Repo

Full repo
  • Spectral Graph Survey-driven entry point to spectral graph theory, graph signals, and modern graph learning. 175
  • Fusion GAN Music generation project on genre fusion with adversarial dual learning. 19
  • XFlow Python library for graph flow simulation, diffusion, and network dynamics experiments. 147
  • SumoXPypsa Coupled transportation-power simulation stack behind the WWW 2026 line of work. 2