I clustered directly on the embeddings (calculated with title + abstract). I think the clustering in my implementation is also not perfect: maybe I luckily got away with it because the broad scope of CHI made them more separable than a more dedicated conference like VIS.

The SOM I used is a "lite-version" from the VIS'24 paper VADIS. Regarding the query, I computed the relevance for each paper (i.e., similarity) to all other papers -- the SOM algorithm takes in the relevance info and the paper embedding matrix to produce the circular projection.

I've been looking forward to your 2025 version! My approach required precomputing everything to generate the circular SOM visualization -- couldn’t support random queries 😆 Yours could handle it!

This wouldn’t be possible without:
• VADIS (VIS’24) for Circular SOM
• GistVis (CHI’25) for in-situ viz
• Project inspiration from Prof. @johnguerra.bsky.social’s CHI’24 Paper Explorer

Proud to build on these amazing works! 🔥

Alternatively, discover in a whole new dimension! ✨
🔍 Holistic view: Explore UMAP/t-SNE projections to see how papers connect.
🏆 Award-winning filter: Find out what made them stand out from the crowd.

Dive in & see the bigger picture!

Explore the papers in a query-based way! 🔍
🎯 Query-based discovery: Pick an anchor paper & instantly map its connections.
📊 Intuitive: Closer (to center & each other) = more related. Spot hidden trends at a glance!

Perfect for curiosity-driven deep dives.