Skip to content

Commit 0146725

Browse files
authored
Merge pull request #64 from BMSguerra/recsys_papers
add LLM bias and Beyon the past recsys papers
2 parents 36ed7db + 65586c9 commit 0146725

File tree

7 files changed

+53
-14
lines changed

7 files changed

+53
-14
lines changed
File renamed without changes.
File renamed without changes.
Lines changed: 41 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,41 @@
1+
---
2+
layout: post
3+
title: "Biases in LLM-Generated Musical Taste Profiles for Recommendation"
4+
date: 2025-09-07 10:00:00 +0200
5+
category: Publication
6+
author: bmassonisguerra
7+
readtime: 4
8+
people:
9+
- bmassonisguerra
10+
- eepure
11+
- mmoussallam
12+
publication_type: conference
13+
publication_title: "Biases in LLM-Generated Musical Taste Profiles for Recommendation"
14+
publication_year: 2025
15+
publication_authors: Bruno Sguerra, Elena V Epure , Harin Lee, Manuel Moussallam
16+
publication_conference: "RecSys"
17+
publication_code: "https://github.com/deezer/recsys25_llm_biases"
18+
publication_preprint: "https://arxiv.org/pdf/2507.16708"
19+
domains:
20+
domains:
21+
- RECSYS
22+
---
23+
24+
We explore a novel use case for Large Language Models (LLMs) in recommendation: generating natural language user taste profiles from listening histories. Unlike traditional opaque embeddings, these profiles are interpretable, editable, and give users greater transparency and control over their personalization. However, it is unclear whether users actually recognize themselves in these profiles, and whether some users or items are systematically better represented than others. Understanding this is crucial for trust, usability, and fairness in LLM-based recommender systems.
25+
26+
<div class="publication-illustration">
27+
<img
28+
style="width: 80%;"
29+
src="{{ '/static/images/publis/sguerra25recsys/overview_LLM_bias.jpg' | prepend: site.url }}"
30+
alt="Overview of the methodology."/>
31+
</div>
32+
33+
To study this, we generate profiles using three different LLMs and evaluate them along two dimensions: self-identification, through a user study with 64 participants, and recommendation performance in a downstream task. We analyze how both are affected by user attributes (e.g., age, taste diversity, mainstreamness) and item features (e.g., genre, country of origin). Our results show that profile quality varies across users and items, and that self-identification and recommendation performance are only weakly correlated. These findings highlight both the promise and the limitations of scrutable, LLM-based profiling in personalized systems.
34+
35+
<div class="publication-illustration">
36+
<img
37+
style="width: 80%;"
38+
src="{{ '/static/images/publis/sguerra25recsys/item_coeff_CR.jpg' | prepend: site.url }}"
39+
alt="Overview of the methodology."/>
40+
</div>
41+

_posts/2025-09-07-recsys-tran.md

Lines changed: 12 additions & 14 deletions
Original file line numberDiff line numberDiff line change
@@ -1,14 +1,10 @@
11
---
22
layout: post
3-
tags:
4-
- recsys
5-
- ACTR
6-
- multimodal
73
title: "'Beyond the past': Leveraging Audio and Human Memory for Sequential Music Recommendation"
8-
date: 2025-09-07 16:00:00 +0200
4+
date: 2025-09-07 10:00:00 +0200
95
category: Publication
106
author: vatran
11-
readtime: 1
7+
readtime: 4
128
people:
139
- vatran
1410
- bmassonisguerra
@@ -18,21 +14,23 @@ people:
1814
publication_type: conference
1915
publication_title: "'Beyond the past': Leveraging Audio and Human Memory for Sequential Music Recommendation"
2016
publication_year: 2025
21-
publication_authors: Viet-Anh Tran, Bruno Sguerra, Gabriel Meseguer-Brocal, Lea Briand, Manuel Moussallam
22-
publication_conference: RecSys
23-
publication_preprint: "https://arxiv.org/pdf/2507.17356"
17+
publication_authors: Viet-Anh Tran, Bruno Sguerra, Gabriel Messeger-brocal, Lea Briand, Manuel Moussallam
18+
publication_conference: "RecSys"
2419
publication_code: "https://github.com/deezer/recsys25-reacta"
20+
publication_preprint: "https://arxiv.org/pdf/2507.17356"
21+
domains:
2522
domains:
2623
- RECSYS
2724
---
2825

29-
On music streaming services, listening sessions are often composed of a balance of familiar and new tracks. Recently, sequential recommender systems have adopted cognitive-informed approaches, such as Adaptive Control of Thought-Rational (ACT-R), to successfully improve the prediction of the most relevant tracks for the next user session. However, one limitation of using a model inspired by human memory (or the past), is that it struggles to recommend new tracks that users have not previously listened to. To bridge this gap, here we propose a model that leverages audio information to predict in advance the ACT-R-like activation of new tracks and incorporates them into the recommendation scoring process. We demonstrate the empirical effectiveness of the proposed model using proprietary data, which we publicly release along with the model's source code to foster future research in this field.
26+
On music streaming services, listening sessions are often composed of a balance of familiar and new tracks. Recently, sequential recommender systems have adopted cognitive-informed approaches, such as Adaptive Control of ThoughtRational (ACT-R), to successfully improve the prediction of the most relevant tracks for the next user session. However, one limitation of using a model inspired by human memory (or the past), is that it struggles to recommend new tracks that users have not previously listened to. To bridge this gap, here we propose a model that leverages audio information to predict in advance the ACT-R-like activation of new tracks and incorporates them into the recommendation scoring process. We demonstrate the empirical effectiveness of the proposed model using proprietary data, which we publicly release along with the models source code to foster future research in this field.
3027

3128
<div class="publication-illustration">
3229
<img
33-
style="width: 75%;"
34-
src="{{ '/static/images/publis/tran25recsys/reacta.png' | prepend: site.url }}"
35-
alt="Architecture of REACTA model (dashed arrows are for inference time)."/>
30+
style="width: 80%;"
31+
src="{{ '/static/images/publis/tran2025recsys/reacta_architecture.jpg' | prepend: site.url }}"
32+
alt="Overview of the methodology."/>
3633
</div>
3734

38-
This paper has been accepted for publication in the proceedings of the 19th ACM Conference on Recommender Systems (RecSys 2025).
35+
36+
2.24 MB
Loading
2.51 MB
Loading
3.51 MB
Loading

0 commit comments

Comments
 (0)