@rpgove
Followers
516
Following
430
Media
136
Statuses
1K
Data vis-à-vis science. Award winning researcher. Rejected by all the top conferences. I've been called the Taylor Swift of data viz.
DC
Joined October 2010
New guidelines for using t-SNE, and evaluation of prior guidelines. And ✨NEW✨: Auto t-SNE, a neural network system to automatically find good perplexity, learning rate, and exaggeration! 📈. Paper: Blog: (1/9).
1
0
5
I'm not very active on Mastodon, but then again I'm not very active on Twitter either
0
0
2
I'm curious how this new Rick & Morty/The Walking Dead crossover series is going to turn out
0
0
0
RT @VizSec: The 2022 Symposium on #DataVisualization for #cybersecurity is only 16 days away. Looking forward to seeing you all on Wednes….
0
5
0
RT @graphicacy_: Our team is looking for Fall and Winter #dataviz engineering interns!. Are you interested in building on and leveraging yo….
0
1
0
0
182
0
RT @syntagmatic: A step-by-step guide to improving a chart design by applying WCAG accessibility principles. Finally publishing some of my….
0
79
0
RT @pickover: Mathematics, computer programming, engineering, motion, creativity. This is a linkage-mechanism for converting binary numbe….
0
392
0
RT @WHCOS: COVID cases are up, so the Biden Admin is making more free COVID tests available -- 8 per household, delivered to your door, fre….
0
1K
0
RT @FrankElavsky: The 5 most important things you can do at a bare bare minimum to make your charts and graphs more accessible:. (Note this….
0
3
0
Coming up in a couple hours, I'm presenting our Auto t-SNE work at Vis Meets AI @PacificVis! Interested in #visualization meets #machinelearning?.Watch the presentation! Or read a sneak peak blog post:
0
0
2
For the #tsne #machinelearning #dataviz users: our new guidelines for using t-SNE can help you get better visualizations!. We'll present this work in 2 weeks at Visualization Meets AI @PacificVis!
New guidelines for using t-SNE, and evaluation of prior guidelines. And ✨NEW✨: Auto t-SNE, a neural network system to automatically find good perplexity, learning rate, and exaggeration! 📈. Paper: Blog: (1/9).
0
0
1
We will present this work on April 11 at Visualization Meets AI, part of @PacificVis. It will appear in a special issue of the Journal of Visual Informatics.(9/9).
0
0
2
In summary:.* Use OpenTSNE.* If not using OpenTSNE, use perplexity=16, learning rate=10, exaggeration=1.* Learning rate = n/12 is also good.* Keep perplexity in 2-16, learning rate in 10-640, and exaggeration in 1-8. Our paper has lots more detail! (8/9).
1
0
1
However, we failed to replicate the recommendation to use perplexity = n/100. This does not appear to generalize from large omics data to smaller tabular data 📉.(7/9).
1
0
0
Our work also replicates some prior research by @Anna_C_Belkina et al. (2019) and @hippopedoid & @CellTypist (2019):.📈 Learning rate = n/12 is good practice and generalizes from large omics data to smaller tabular data. (6/9).
1
0
2
Importantly, we find room for future work: an "oracle" that always picks the best combo from a grid of t-SNE parameters outperforms the other methods. This leaves room to improve t-SNE guidelines. (5/9).
1
0
1
We evaluate our neural network, our empirical parameter combo, and the OpenTSNE defaults. They all perform comparably and better than both the sklearn defaults and a mix of other guidelines. Note: our neural network is unoptimized, so better performance is possible. (4/9)
1
0
0
We also developed Auto t-SNE to automatically find good t-SNE parameters! It uses a neural network to predict t-SNE accuracy given a data set and parameter combo. Picking the best parameters takes ~1s, way faster than a brute force approach! Similar to work by @OhHyunKwon.(3/9)
1
0
2