Language Model-Based Playlist Generation Recommender

This repository contains the implementation of a novel approach to playlist generation using language models. Our method leverages the thematic coherence between playlist titles and their tracks by creating semantic clusters from text embeddings. We fine-tune a transformer model on these clusters to generate playlists based on cosine similarity scores between known and unknown titles, utilizing a voting mechanism for final recommendations.

The repository includes code for preprocessing, model training, evaluation, and generating recommendations. For more detail, please refer to the paper.

Related links - Online demo - code - Zenodo repository including the best trained model.

1. Transform and pre-process the dataset

Run:

bash python3 transform-dataset/json2csv.py

to convert the JSON slices of the dataset into user-friendly CSV files:

2. Embedding generation and clustering

First, playlists titles and tracks are embedded using a pre-trained SentenceBERT model and stored in a 'pickle' file:

bash python3 clustering-no-split/embeddings/track_embeddings_no-split.py

Then, the K-means clustering algorithm is applied to create the clusters, and the generated 'csv' file is modified to calculate and include the percentage of exact matches:

bash python3 clustering-no-split/clusters/clustering-no-split.py clustering-no-split/clusters/percent-no-split.py

Apply the clean algorithm to remove miscellaneous clusters:

bash python3 clustering-no-split/clean/clean.py

Finally, randomly split the clusters, ensuring a representation of each cluster in both train, test and validation sets:

bash python3 clustering-no-split/split/split.py

3. Finetuning

Train the SentenceBERT model with two loss functions (cross-entropy and triplet loss) to better capture thematic similarities:

bash python3 finetuning/cross_entropy_model_finetuning.py finetuning/finetuning_triplet_loss.py

4. Generate the embeddings for playlists titles using the fine-tuned models

Run :

bash python3 embeddings/playlists_embeddings_final.py

to generate embeddings for playlist titles using the fine-tuned models.

Make sure to adjust the model path to select either the triplet loss model, the cross-entropy loss model or the pretrained model.

5. Generate the recommendations and evaluate the models

Evaluate the metrics for a given test playlist:

bash python3 similarity/test_1_playlist_finetuned_model.py

Generate the recommendation for a playlist title:

bash python3 similarity/recommend.py

Assess the model’s overall performance on the complete test set:

bash python3 similarity/testset_test_model.py

Make sure to adjust the model path to select either the triplet loss model, the cross-entropy loss model or the pretrained model.

Citation

If you use this software, please cite (bib file):

Enzo Charolois–Pasqua, Eléa Vellard, Youssra Rebboud, Pasquale Lisena, and Raphael Troncy. 2025. A Language Model-Based Playlist Generation Recommender System. In Proceedings of the Nineteenth ACM Conference on Recommender Systems (RecSys ’25), September 22–26, 2025, Prague, Czech Re- public. ACM, New York, NY, USA, 11 pages. https://doi.org/10.1145/3705328.3748053