Multimodal LightGCN

This project enhances LightGCN with multimodal embeddings.

Various architectures of fusion of collaborative and multimodal embeddings are tested.

The results show that multimodal information is most effective in settings where collaborative information is scarce.

Running

This project uses the uv package and project manager.

src/config.toml contains the configuration for the model. Attributes and possible values:

• fusion_modalities: concat|mean|sum|max|min|prod
• fusion_type: late|early|inner
• weighting: false|alpha|normalized|equal

To run, run src/main.py; configuration will be loaded from src/config.toml; the GPU with the lowest used memory will be used.

logs/ contains the logs of the training with the metrics for each epoch; the first line contains a JSON representation of the configuration used to produce that file.

src/graph.ipynb contains utilities to read and filter the logs, show them in tabular format and produce graphs.

src/grid_search.py lets you run all experiments in sequence for a single dataset.