[Seeking Avenues] Building a Music Recommendation System

[foggy relic]

Hello everyone!

I'm diving into an exciting project and could really use your collective wisdom. I want to develop a music recommendation system but am at square one regarding where to begin. With so many tools and technologies available, the starting point feels overwhelming, I feel completely hamstrung.

I've heard a lot about Large Language Models (LLMs) and their capabilities in enhancing recommendation systems through natural language understanding and user interaction. However, I'm not sure if this is the best path for a music recommendation system, or if there are more specific technologies or frameworks that would better serve this purpose.

Additionally, I'm curious about any recent open-source projects in this area. If there are existing projects that I could contribute to or use as a foundation, that could accelerate my learning and development process. Any pointers towards GitHub repositories or communities would be incredibly helpful.

Lastly, if anyone has a roadmap or resources that outline the steps to create a music recommendation system — from data collection and processing to algorithm implementation and user interface design — I would be immensely grateful. For now I did even not have the data. I think I will create my own dataset from Youtube Music because I want to create a recommendation system from my musical tastes. I also must find a way to convert my songs to embeddings. I would like my recommendation-system be more focused on electronic dance music (EDM).

Thank you in advance for your guidance and recommendations. Looking forward to learning from and possibly collaborating with this knowledgeable community!

Best,
William.

[pliant wolf]

A recommendation system doesn't need any sort of large models.

https://pytorch.org/blog/introducing-torchrec/

Here's a very large guide for RS kaggle notebook. You'll get an understanding of how it works

https://www.kaggle.com/code/prashant111/recommender-systems-in-python

[foggy relic]

Thank you very much for your answer. Do the RS take into account the sequential aspects? In my RS what really matters to me is to generate a consistent sequence of songs, with appropriate transitions between the songs. Should I dig a particular domain to success to implement that?

[karmic oriole]

Recommendation systems can really be thought of as bipartite graph matching. There is P a set of products : P = {product 1,..., product n}

There is U a set of users : U = {user 1, ..., user m}

It's then about finding the best links from U to P (both ways) that will maximize a certain function

Anyway my point is that it can be formulated as a graph problem and hence graph neural networks can be a good option (maybe not a good start)

[round lodge]

Off the cuff, you can start with just a bunch of tags for songs (artist, album, genre, etc). See how that does.

Then you can use vector embeddings with models like HuBERT to generate embeddings and use KNN to get the embeddings of similar songs. Compare that to the first system.

Then you can stack the two: stage 1 reduced the search space by pulling top recommendations based on song tags and stage 2 takes that short list and puts out which songs have better embeddings match. Again, see how that goes. It’s very similar to ReRank search for RAG systems.

At no point do you really “need” anything LLM related at all.

[foggy relic]

Thank you for your answers.

@round lodge your roadmap is really good ! I would also like generate playlists/tracklists (the best sequences of songs) from the recommended songs? How should I proceed? Can one say that it is to find a path in the recommended songs?