The Pitfalls of Using ML-Based Optimization - IPAM at UCLA

Explore a conference talk that delves into the limitations of machine learning-based optimization in solving graph problems. Examine two specific cases where traditional optimization methods outperform ML approaches: the Force Path Cut problem and the Hypergraph Discovery problem. Gain insights into the speaker's conjectures about why ML-based optimization struggles with these challenges. Learn about the implications for targeted diffusion, cyber security, and the potential instability of embeddings. Discover the importance of considering alternative optimization techniques when dealing with complex graph structures and energy-constrained agent-task allocation scenarios.

Introduction
Instability of embeddings
Degenerate core
All empirical
Point of instability
Negative Sampling
Topological Shortcuts
Annotation and Balance
Man in the Middle Attacks
Force Path Cut
Weighted Graphs
MLbased Optimization
Application
Targeted Diffusion
Attack Vector
Budget
Diffusion
Cyber Security