Explanation

In a margin classifier—most famously the Support Vector Machine—you don’t just look for any separating hyperplane, you look for the one that maximizes the margin, i.e. the perpendicular distance from the hyperplane to the closest training point of either class. The hyperplane in a linear model is defined by the equation w^Tx + b = 0, where w is the normal vector that points orthogonal to the decision surface and b is the bias term that shifts the plane. The margin is mathematically 1/‖w‖ (distance to the nearest point) and the distance between the two parallel “supporting” hyperplanes w^Tx + b = ±1 is 2/‖w‖. By solving a convex optimization problem that maximizes this distance, the classifier becomes robust to small perturbations in the data.

Let’s illustrate with a tiny 2‑D dataset:

Class +1: (2, 2) and (2, 0)
Class –1: (0, 0) and (0, 2)

Because the points of each class share the same x‑coordinate, a vertical line perfectly separates them. The SVM constraints are

$$ \begin{aligned} w^{\top}x_i + b &\ge 1 \quad \text{for } y_i = +1,\\ w^{\top}x_i + b &\le -1 \quad \text{for } y_i = -1. \end{aligned} $$

Plugging the four points into these inequalities yields the simple solution w = [1, 0] and b = –1, so the optimal hyperplane is

$$ x_1 - 1 = 0, $$

i.e. the line x = 1. The distance from this line to any of the four points is

$$ \frac{|w^{\top}x + b|}{\|w\|} = \frac{|0 - 1|}{1} = 1, $$

so the margin (distance to the nearest point) is 1. The two supporting hyperplanes are x = 0 (w^Tx + b = –1) and x = 2 (w^Tx + b = +1), and their separation is 2, confirming the margin of 1.

Naive Response

A margin classifier is a linear model that tries to separate the two classes with a line (or hyperplane) and wants the line to be as far away as possible from the points. The hyperplane is given by wᵀx + b = 0. For a tiny dataset you can just draw a line between the groups and measure the shortest distance to any point as the margin.

Good Response

First, I would define a margin classifier as a linear model that not only separates the classes but does so by maximizing the geometric margin—the smallest perpendicular distance from the decision boundary to any training example. The decision boundary is a hyperplane defined by wᵀx + b = 0, where w is orthogonal to the plane. In the SVM formulation we enforce the constraints wᵀx_i + b ≥ 1 for positive examples and wᵀx_i + b ≤ –1 for negative examples; the margin then equals 1/‖w‖.

Using the provided four‑point example, I write the constraints explicitly:

$$ \begin{aligned} 2w_1 + 2w_2 + b &\ge 1,\\ 2w_1 + 0w_2 + b &\ge 1,\\ 0w_1 + 0w_2 + b &\le -1,\\ 0w_1 + 2w_2 + b &\le -1. \end{aligned} $$

Because the points share the same x‑coordinate within each class, the optimal solution is w = [1, 0] and b = –1, giving the hyperplane x₁ – 1 = 0 (a vertical line at x = 1). The norm ‖w‖ = 1, so the margin is 1/‖w‖ = 1. The distance from the hyperplane to any of the four points is |wᵀx + b|/‖w‖ = 1, confirming that the nearest points (the support vectors) lie exactly one unit away. The two parallel supporting hyperplanes are x₁ = 0 and x₁ = 2, whose separation is 2, which matches the 2/‖w‖ formula.

Finally, I would emphasize why this matters: maximizing the margin yields a classifier that is less sensitive to noise and typically generalizes better, which is the core intuition behind SVMs.

Assessment

The naive answer mentions the right ingredients—hyperplane, margin, distance—but stays at a high level and never shows how to turn the intuition into a concrete solution. The good response walks through the full reasoning: it states the formal definition, writes the SVM constraints, solves for w and b, derives the hyperplane equation, computes the margin using the ‖w‖ formula, and connects the result back to the supporting hyperplanes and the generalization benefit of a large margin. This depth demonstrates both mathematical rigor and practical insight, which is exactly what interviewers look for.

150. Hyperplane Margin Classifier

Similar Questions