Very cool! I’ve done visualizations on similar size datasets by mapping tsne output to fake latitude/longitude coordinates and rendering it with mapbox gl. It feels like a hack, but then you get all of the panning/zooming interaction, styling, and rendering labels with collision avoidance for free. For smaller datasets I pass in geojson directly, but for larger ones I precompute vector tiles so you can lazy-load more of the data when you zoom in. There are a lot of things built for mapping that are relevant for other domains, I wish that there was a common base that could be extracted that mapping libraries adapt to geographic data but data visualizations could adapt to non-geographic data.<p>I’ll have to play around with your technique more, encouraging to see you get good performance eagerly loading a million data points!