PCA is just a transformation of your feature space via centering and rotation such that your components (the resulting basis vectors) are pointing in the direction of greatest variance in descending order. That is component 1 explains the most variance, component 2 the next, and the last explains the least. So this is really just a mapping from an n-dimensional space to an n-dimensional space.
It is common to then choose only a subset of the space to perform your analysis. By default the Classification Learner App will choose enough components to explain at least 95% of the original variance. In your case it seems you had 4 components and more than 95% of the variance in that 4-dimensional space can be explained by a 3-dimensional subspace. However, you should not think of this in terms of the original 4-factors as there is some information present from all of those 4-dimensions (excepting in the special case of multi-collinearity).
