Visualization Assessment: A Machine Learning Approach

Abstract

Researchers assess visualizations from multiple aspects, such as aesthetics, memorability, engagement, and efficiency. However, these assessments are mostly carried out through user studies. There is a lack of automatic visualization assessment approaches, which hinders further applications like visualization recommendation, indexing, and generation. In this paper, we propose automating the visualization assessment process with modern machine learning approaches. We utilize a semi-supervised learning method, which first employs Variational Autoencoder (VAE) to learn effective features from visualizations, subsequently training machine learning models for different assessment tasks. Then, we can automatically assess new visualization images by predicting their scores or rankings with the trained model. To evaluate our method, we run two different assessment tasks, namely, aesthetics and memorability, on different visualization datasets. Experiments show that our method can learn effective visual features and achieves good performance on these assessment tasks.