In articles like these:
the authors mention using the dimension reduction vectors as features for a classifier (be it RF or XGB).
However, the original t-SNE application is fully heuristic in nature. I can imagine performing t-SNE—>classifier on the training data. But, there isn’t a way to apply a “trained t-SNE” model on the test data.
So, my question is: How does one apply t-SNE in a classification scheme? What am I missing?