It is the 1990s, and two point clouds sit in thirty-dimensional space, perfectly separable by a hyperplane — but infinitely many hyperplanes separate them, and all score zero errors on the data you have. Zero training error is not the goal; the goal is the future. Find the classifier that best guarantees low error on examples not yet seen, which means reasoning about the margin — the width of the buffer between the classes — rather than merely fitting the training points. The generalization guarantee is a theorem, not an empirical hope. Get it wrong and you ship a boundary that aces the test set and fails in the wild, while looking perfect on paper — fitting the past is trivial; the margin buys the future.
Chose The network census — wrong. Graph representation learning was the one that fit.
The scene is the 1990s, when Leskovec was a student rather than a contributor, and support vector machines reached him as settled technology. They appear throughout his early data-mining work as the standard classifier sitting downstream of whatever features or embeddings he had constructed, and he teaches the margin intuition, but the generalization theory the problem actually asks for, VC dimension, margin bounds, structural risk minimization, is entirely inherited. There is also a mismatch of instinct. His characteristic response to a hard classification problem is not to reason about the geometry of a fixed thirty-dimensional feature space; it is to change the representation until the problem becomes easy, which is what node2vec (2016) and GraphSAGE (2017) do. That reflex would serve him well in practice here while leaving the theorem, the part this problem scores, unwritten.
Urtasun's research sits in 3D computer vision and autonomous driving: structured prediction for scene understanding, efficient inference in Markov random fields, KITTI-era benchmarks, and later self-driving stacks at Uber ATG and Waabi. She came up when max-margin methods were live infrastructure — structured SVMs and max-margin Markov formulations were standard tools in her community, and some of her early work used latent structured max-margin training — so the margin principle is genuinely part of her working toolkit, extended to combinatorial output spaces. But her extensions concern structured outputs, latent variables, and perception pipelines; two separable clouds with a scalar label exercise none of it, and the core distribution-free guarantee is inherited from the theorists she cites. Practical fluency with the winning family, no claim on its founding theorem, and a problem too clean for her machinery.
Battle #100 · 8/10/2026, 11:37:52 AM · this result is deterministic: the same two personas on this problem always resolve the same way.