In nearest-neighbor un-embedding, the signed distances calculated are used to determine which class a new data point belongs to, and the closest class vector is always the correct classification.

Nearest‑neighbor un‑embedding works by taking a new data point, computing its signed distance to each class vector, and then picking the class whose vector is closest. The signed distance tells us how far the point is from a class direction, positive if it lies on the same side of the decision boundary and negative otherwise. Because the distance is signed, a smaller absolute value means the point is nearer to that class’s direction in the embedding space. In practice, the class with the smallest signed distance is taken as the predicted label, and this rule is assumed to give the correct classification. For example, if a point has distances –0.