Disentangled Representation of Data Distributions in Scatterplots

We present a data-driven approach to obtain a disentangled and interpretable representation that can characterize bivariate data distributions of scatterplots. We first collect tabular datasets from the Web and build a training corpus consisting of over one million scatterplot images. Then, we train a state-of-the-art disentangling model, β-variational autoencoder, to derive a disentangled representation of the scatterplot images. The main output of this work is a list of 32 representative features that can capture the underlying structures of bivariate data distributions. Through latent traversals, we seek for high-level semantics of the features and compare them to previous human-derived concepts such as scagnostics measures. Finally, using the 32 features as an input, we build a simple neural network to predict the perceptual distances between scatterplots that were previously scored by human annotators. We found Pearson's correlation coefficient between the predicted and perceptual distances was above 0.75, which indicates the effectiveness of our representation in the quantitative characterization of scatterplots.