1 of 19

A Language Model-Based Playlist Generation�Recommender System

Enzo Charolois–Pasqua · Eléa Vellard · Youssra Rebboud · Pasquale Lisena · Raphael Troncy

2 of 19

Love

Christmas

Hype

Country

2

Artist/Genre/Style (25%)

Cunningham, Sally Jo; Bainbridge, David; Falconer, Annette�"More of an art than a science": Supporting the creation of playlists and mixes. ISMIR 2006

Event or Activity (25%)

Romance (19%)

Mood (16%)

Other (2.6%)

3 of 19

Playlist names are not necessarily bound to a category

  • Total freedom of expression for the user
  • How to automatically handle them?�(e.g. cold start recommendation?)

“Housewarming Party”

3

“Country summer”

“Spring awakening”

(real playlist names from the Million Playlist Dataset)

“Marina and the diamonds”

“Dancing in the kitchen”

4 of 19

4

What are the best tracks for this playlist title? 🍋

5 of 19

Problem statement

  • Goal:�Generate a playlist using only its title
  • Impact:
    • Cold-start recommendation
    • Automatic playlist generation
  • Challenges:
    • Decontextualized natural language
    • No dependency on a set lexicon
    • High variability in playlist titles

5

  • Million Playlist Dataset
    • Initially released by Spotify for �the 2018 RecSys Challenge
    • 1 million playlists
    • 2+ million tracks

Ching-Wei Chen, Paul Lamere, Markus Schedl, and Hamed Zamani. �Recsys Challenge 2018: Automatic Music Playlist Continuation. RecSys 2018

6 of 19

Related work

Text & NLP Approaches

  • Playlist creation & categories (Cunningham et al., 2006)
  • Classification with topic models (Fields et al., 2010)
  • Songs as token sequences (McFee & Lanckriet, 2011)
  • Emotion-based playlist recs (Nair et al., 2021)
  • Metadata & knowledge graphs (Gabbolini & Bridge, 2023)

Large Language Models in Music Recs

  • Playlist title generation (Doh et al., 2021, Kim et al., 2023)
  • Text2Playlist (Delcluze et al., 2025)
  • Text2Tracks (Palumbo, 2025) & TALKPLAY (Doh et al., 2025)

6

Gap: Limited focus on playlist titles as input

Playlist Titles in Recommender Systems

RecSys Challenge 2018: cold-start with titles

  • Monti et al., 2018 – RNN embeddings
  • Faggioli et al., 2018 – similarity matrices
  • Kim et al., 2018 – LSTM with n-grams

7 of 19

Approach

#clustering #fine-tuning #language-models #semantic-similarity

7

Phase 1

Fine-Tune a Language Model

Phase 2

Predict with Semantic Similarity

8 of 19

Main idea

8

words occurring often together in sentences

close in the�embedding space

titles occurring often together in playlists

close in the�embedding space

9 of 19

Phase 1: Fine-tune a Language Model

  • Sentence-BERT embeddings
  • Semantic clustering K-Means, k = 15
  • Remove miscellaneous clusters
  • Fine-tuning all-MiniLM-L6-v2
    • Cross-entropy loss (best), Triplet loss
    • 15 epochs (early stopping), LR 2 e-5

9

MPD

Pretrained�SBERT

Playlist embeddings

Clustering

Fine-tuning

Fine-tuned language

model

10 of 19

Phase 2: Predict with Semantic Similarity

  • Relevant playlist retrieved using nearest-neighbor from Gensim
  • Voting system: Count track frequency in top-50 neighbors, return the most frequent tracks

10

Fine-tuned language

model

New�Playlist Title

MPD

Embedding representation

Embedding representation

K most relevant playlists

Ranking based on cosine similarity

N recommended tracks

Voting mechanism

11 of 19

Evaluation

#quantitative #qualitative

11

12 of 19

Quantitative evaluation

12

Method

R-Precision

NDCG

Monti et al. (Only title)

0.0837

0.1260

Faggioli et al. (Only title)

0.1093

0.2451

Kim et al. (Only title)

0.0760

0.1866

Pre-trained

0.1570

0.2731

Fine-tuned (cross-entropy)

0.1556

0.2825

Fine-tuned (triplet loss)

0.1285

0.2297

from RecSys Challenge 2018

our method

13 of 19

Qualitative evaluation

Rock Classics

13

▶ My Heart Will Go On� Céline Dion

▶ Highway to Hell� AC/DC

▶ Smells Like Teen Spirit� Nirvana

▶ It’s My Life� Bon Jovi

14 of 19

Qualitative evaluation

  • On a subset of 22 selected playlists
  • Human Judgment of Playlist Relevance
  • Qualitative score = nb. valid tracks

total recommended tracks

14

Method

Quality@10

Quality@66

Pre-trained

0.7376

0.7231

Fine-tuned (cross-entropy)

0.7789

0.7719

Fine-tuned (triplet loss)

0.7533

0.7461

15 of 19

Predict with LLMs

#prompt #zero-shot #few-shot

15

16 of 19

LLM Generation

You are an expert in music playlist generation.

Your task is to generate the continuation of a playlist given only its title and five example songs with their artists.

Important:

• You have to select only songs released before October 2017.

• Propose a COMPLETE playlist consisting of exactly 10 songs.

• All recommended songs must be UNIQUE and must not repeat any of the five example songs provided.

Playlist Title: "{playlist_title}"

Examples:

(1) {"song": "{song1}", "artist": "{artist1}"}

(2) . . .

Output format (strict):

[

{"song": "<title>", "artist": "<artist>"},

...

]

Answer ONLY with the JSON list exactly as specified above. Do not output anything else.

16

Setting a persona

Give precise instructions

Include N samples

Ask a specific output

approximation to match MPD content

17 of 19

Evaluation of the LLM

17

Metric

GPT-4o�(0-shot)

GPT-4o�(5-shot)

Our Method (FT-C)

Precision@10

0.0636

0.1227

0.1793

Recall@10

0.0073

0.0197

0.0382

MRR@10

0.1636

0.2505

0.3254

R-Precision@10

0.0157

0.0338

0.0496

NDCG@10

0.1900

0.3249

0.3740

Qualitative Score

0.7175

0.7953

0.7789

  • Also tried Llama and Zephyr (results in the paper)
  • GPT-4o improves from 0-shot → 5-shot, but still falls short on retrieval metrics.

18 of 19

Conclusion

Contributions

  • Pipeline for playlist generation from titles only
  • Outperformed state-of-the-art title-based methods
  • First assessment of prompt-based LLM

Future Work

  • Mitigate popularity bias �→ integrate diversity & novelty metrics
  • Playlist continuation�(use first N tracks as seeds)
  • Hybrid systems�→ title embeddings + collaborative filtering

18

19 of 19

Thank you!

This presentation

bit.ly/lmrec-2025

WE HAVE A DEMO!

playlist-recommendation.tools.eurecom.fr

SPOT #21 - Thu Sept, 25 - Poster Session