Apps · Concepts
Pattern Classification Lab
Every classifier divides feature space into regions. Place points, pick a model and watch the boundary form. Then add noise, overlap or a tricky shape and see which models cope.
Decision regionsA ●B ■C ▲hollow = test points
Training losscross-entropy per epoch
What each model can draw
| Model | Boundary shape | Watch for |
|---|---|---|
| k-nearest neighbours | follows the data locally | k = 1 wraps around every noisy point; large k blurs small groups away |
| Logistic regression | straight lines only | cannot solve moons, rings, XOR or spirals at all |
| Decision tree | axis-aligned boxes | deep trees carve tiny boxes around noise |
| Random Forest | averaged boxes, smoother | still stair-stepped on diagonal boundaries |
| Neural network | any smooth curve, learned by gradient descent | too few units underfit; a learning rate that is too high makes the loss jump |
Fault diagnosis works the same way in more dimensions: each machine window is a point, each feature an axis, and the model's regions decide the diagnosis. The Feature Space Explorer shows real vibration features in this kind of view. Read more in Fault classification.