Predicting User Preferences of Dimensionality Reduction Embedding Quality

Abstract

Fig. 1: Overview of the best embeddings, as predicted by our method, for three of the datasets we have collected: MNIST, photos of flowers, and paintings. On the left of each sub-figure, we have an embedding of embeddings (called a metamap) where each square represents an embedding that was considered in our study and the contour color coding represents goodness of embedding (with dark blue tones meaning "good" and dark red tones meaning "bad"). On the right of the sub-figures, we have the top 3 best embeddings. The background of the metamap is visualized using a "goodness" score for the embeddings outputted from our model. The top 3 MNIST embeddings (blue background -high score) are of higher quality than the top 3 painting embeddings (red background -low score).