Support-Vector Networks

Why this one

Read this after The Matrix Calculus You Need For Deep Learning if your instinct is to explain learning through gradients, and before Random Forests if you want a clean contrast with ensemble pragmatism. The durable idea here is that classification can be posed as geometry: find a separating surface with the widest safe gap, and let only the borderline examples matter. People often remember the kernel trick as a clever way to get nonlinear models without building features. That is true, but too small. The real lesson is that a model can generalize because the training problem encodes a preference for simple geometry, not because it has fewer parameters or because it guesses the right curve. If you build ranking, retrieval, moderation, or anomaly systems, this paper gives you a mental tool you will reuse: look for the few examples that define the boundary, then ask what space would make that boundary simple.