Towards a comprehensive visualization of structure in data

Dimensional data reduction methods are fundamental to explore and visualize large data sets. Basic requirements for unsupervised data exploration are simplicity, flexibility and scalability. However, current methods show complex parameterizations and strong computational limitations when exploring large data structures across scales. Here, we focus on the t-SNE algorithm and show that a simplified parameter setup with a single control parameter, namely the perplexity, can effectively balance local and global data structure visualization. We also designed a chunk\&mix protocol to efficiently parallelize t-SNE and explore data structure across a much wide range of scales than currently available. Our parallel version of the BH-tSNE, namely pt-SNE, converges to good global embedding, comparable to state-of-the-art solutions, though the chunk\&mix protocol adds little noise and decreases the accuracy at the local scale. Nonetheless, we show that simple post-processing can efficiently restore local scale visualization, without any loss of precision at the global scales. We expect the same approach to apply to faster embedding algorithms other than BH-tSNE, like FIt-SNE or UMAP, thus, extending the state-of-the-art and leading to more comprehensive data structure visualization and analysis.