A Multi-Source Retrieve-Fuse-Rerank System for Conversational Music Recommendation
TL;DR
The ablation study shows that the contextual and utterance dense retrievers provide the largest gains in candidate coverage, while structural retrieval also improves the final ranking, and the benefit of combining complementary retrieval signals with a learned reranker for conversational music recommendation.
Abstract
Conversational music recommendation requires a system to understand user preferences across multiple dialogue turns, retrieve relevant tracks from a large catalog, and generate a suitable response. We present a multi-source retrieve–fuse–rerank–generate pipeline for the RecSys Challenge 2026. The system combines lexical, dense, structural, and visual retrievers to build a diverse candidate pool. A LightGBM LambdaRank model then reranks the candidates using retrieval, metadata, and dialogue-history features. Finally, Qwen3-14B generates a grounded response while keeping the ranked track list fixed. The final submission of Team soussou ranked 11th overall and fifth among academic teams on the official Blind-B leaderboard, with an nDCG@20 of 0.4277 and a composite score of 0.5303. The ablation study shows that the contextual and utterance dense retrievers provide the largest gains in candidate coverage, while structural retrieval also improves the final ranking. These results show the benefit of combining complementary retrieval signals with a learned reranker for conversational music recommendation.