Can it learn with less?
Sparse graphs. Useful features. No signal sacrificed.
Graduate Students Orientation · 2026
Dr. MudassirShabbir
Associate Professor of Computer Science
Syed Babar Ali School of Science & Engineering
Lahore University of Management Sciences
Graph machine learningResilient networked systemsDiscrete geometry
Graphs · Algorithms · Control · Learning
MLControlGeometryNetworks
Graduate orientation · LUMS · 2026
I remove edges
for a living.*
* Very carefully.
Graph learning, resilient networks, and the mathematics hiding underneath.
learn
control
survive
My research asks graphs three awkward questions
Less. Lies. Lookalikes.
✓✓!✓✓“trust me”
Can it survive a liar?
Robots cooperate even when one teammate goes rogue.
≟
Can it prove who it is?
Different graphs can share the same spectral “sound.”
Less · Graph machine learning
Marie Kondo,
but for graphs.
Keep the connections that spark signal.
raw network 28 edges
zero forcing→
learning backbone 11 edges
GraphSAGE
68.7
73.9
+ rank encodingLies · Resilient multi-agent systems
One robot lies.
The team still learns.
“Fly left!”Byzantine agent
Consensus: fly right →
corecritical links protected
2025 · Multi-UAV systemsResilient distributed learningAggregate robustly under adversarial information ↗
2026 · AutomaticaControllability backboneProtect a small core; preserve the ability to steer ↗
Byzantine = a teammate who is confidently wrong on purpose.
Distributed systems cannot simply average every message when an agent may be malicious. The multi-UAV work studies resilient learning under adversarial information. The controllability-backbone paper asks a structural companion question: which minimum set of links must survive so that the same leaders retain at least the original controllability bound?Lookalikes · Spectral graph theory
Graph A: “I’m unique.”
Graph B: “Sure.”
A
same spectrum≠same graph
B
walk matrix
1 0 3 7
0 2 4 9
1 4 8 16
+ arithmetic fingerprint0 2 4 9
1 4 8 16
Your move
Bring me a
messy network.
We’ll find the structure that makes it…
learn with lesssurvive bad actorsprove its identity
?
MLgraphscontrolrobotsalgebrageometry