Improving Neural Network-based Multidimensional Projections

Dimensionality reduction methods are often used to explore multidimensional data in data science and information visualization. Techniques of the SNE-class, such as t-SNE, have become the standard for data exploration due to their good visual cluster separation, but are computationally expensive and don’t have out-of-sample capability by default. Recently, a neural network-based technique was proposed, which adds out-of-sample capability to t-SNE with good results, but with the disavantage of introducing some diffusion of the points in the result. In this paper we evaluate many neural network-tuning strategies to improve the results of this technique. We show that a careful selection of network architecture, loss function and data augmentation strategy can improve results.