Show HN: A navigable map and recommender for 17M music entities

Show HN: A navigable map and recommender for 17M music entities(toposonico.com)

4 points by deppep 16 days ago | 2 comments

Hello HN,

This is toposonico, a music recommender and navigable map. At core it's a skipgram word2vec model trained over ~6M playlists. Tracks are embedded in a 128d space. Embeddings for albums, artists and labels are computed marginalizing over tracks. The 2D map was built with UMAP.

Both the model and UMAP were trained in the cloud over a NVIDIA A100. All things considered it cost me around ~50EUR, over two main training sessions and a few experiments. For the slippy map I experimented with a few libraries. Ended up with Maplibre GL JS. Loved working with it, kudos to their developers. For the recommender indexes I used FAISS, another fantastic piece of software. Pretty happy with the thing running on a small and cheap box.

Two things influenced me in making this. The first is decade-old idea: human navigation and exploration skills work in information spaces too. Many ML concepts fit this idea especially well. It would be nice to see more experiments in this direction. The second goes more like a story. Before moving out and selling my turntable, I used to visit record fairs and I always ended up finding something novel there. To find a new record you didn't sit down and listen to ten records in a row, selected based on a supposed model of your personality. You wandered around, speaking with dealers and looking through the crates they brought with them. The crates often leaned on some genre more than another, reflecting the dealer's history and taste. There were huge stalls and there were small ones. Some were crap. Finding oddities was very easy. I wanted to write something that felt like that.

Let me know what you think, if you find any bugs or if you have any idea for improving it.

Repo link: https://github.com/peppedilillo/toposonico

dangravell 12 days ago |

At the back end of the 90s I did some work with self-organizing maps - collapsing multi dimensional space into a 2D map. I think it's interesting as a data visualisation approach. This looks similar but I'm totally out of date with where SOM went.

When you say 128d space, what are the data that are represented, exactly?

deppep 11 days ago | |

Hello! I find these fascinating too. The 128 dimensions I mentioned are learned by the NN, they are latent space and their actual meaning is ultimately unknown. The number itself is somewhat arbitrary as it was selected optimizing memory usage vs data coverage so that the track embeddings could sit on an A100.