Discogs Rec Model

This repository houses the code related for training the discogs-rec model. It utilizes Spotify's Annoy library and is trained on data from the Discogs data dumps, as well as additional data such as user wants, haves, and pricing collected via webscraping. At the moment Discogs Rec only supports recommendations for electronic music and hip-hop listed on Discogs. If you would like to use the model via API, please see this repo here: discogs-rec-api. If you can't be bothered to setup the API I provided a demo script in the /demo folder which will allow you to get predictions from a discogs URL right away (see example below).

Repository Structure

• discogs_rec/: Contains the source code for the creation of the Recommender System. This includes the preprocessing of the features used to build the model, generating the .ann file, and creating the mappings between release ids and artist/track titles.
• data/: Directory for the training data, annoy index, config, and mappings (gets created when running the container.).
• demo/: Directory where the demo script lives to fetch recommendations.
• tests/: Unit tests.

Setup

1. Clone the repository:

git clone https://github.com/justinpakzad/discogs-rec-model

2. Install dependencies (only necessary if you want to try the demo, otherwise Docker handles it)

uv sync

Docker Commands

To build and run the discogs_rec_model services with docker compose, use the following commands:

docker compose up discogs_rec_model -d

Or to specify a subset of features to be used:

docker compose run --build discogs_rec_model python discogs_rec/main.py --features low median high countries styles wants haves

To run tests:

docker compose run --rm tests

Demo Example

To get recommendations provide a url and use the following command:

uv run demo/demo.py --url "https://www.discogs.com/release/335130-FL-Untitled"

Returns:

[
    {
        "artist": "Subversiv' Effigies",
        "title": "Strictly Unreal",
        "label": "Global Cuts",
        "url": "https://www.discogs.com/release/127580"
    },
    {
        "artist": "Van Chung",
        "title": "Electric Flying",
        "label": "Smash Productions",
        "url": "https://www.discogs.com/release/2132373"
    },
    {
        "artist": "DJ 4T4",
        "title": "Last Nite / It's Just Begun",
        "label": "Swet",
        "url": "https://www.discogs.com/release/731782"
    },
    {
        "artist": "Jack LeHit And The Sample Minds",
        "title": "The Haunted House Of Soul",
        "label": "Sound Of Belgium",
        "url": "https://www.discogs.com/release/618792"
    },
    {
        "artist": "Common Factor",
        "title": "Motion Transition EP",
        "label": "SSR (4)",
        "url": "https://www.discogs.com/release/38842"
    }
]

Notes

Feel free to experiment with the number of features used to generate the Annoy index file. I've experimented with using slightly fewer features and found it resulted in more random but interesting recommendations. You can also adjust the weight of the feature matrices for varied results in weights_dict which resides in the preprocessing.py file. Lastly, tweaking the number of trees (n_trees) in the Annoy file can enhance recommendation accuracy at the cost of a larger .ann file and increased memory usage during queries, as found in utils.py. If you are having issues running this locally due to a compute issue, consider lowering the number of trees. This is an ongoing project, I will be trying to add/update new features and fix any bugs that may appear along the way.

Contributing

Any contributions are welcome, please feel free to submit a Pull Request.
How to contribute:

1. Fork the repository
2. Create a feature branch
3. Commit your changes
4. Push to the branch
5. Open a Pull Request