Beyond Single-Signal Retrieval: A Graph-Aware Agentic Framework for Conversational Music Recommendation
TL;DR
Experimental results show the pipeline outperforms baselines in ranking, diversity, and explanations, proving that combining graph embeddings with agentic reasoning successfully bridges long-term profiles and immediate intent.
Abstract
Conversational music recommendation requires fusing multi-turn dialogue with multimodal item representations to capture both explicit requests and implicit context. This task is notoriously difficult due to vague user intents and preference signals scattered across turns. In this paper, our team, Poliba T1 b, introduces a novel graph-aware agentic framework to tackle the RecSys Challenge 2026. Our approach extracts long-term collaborative preferences via a heterogeneous Graph Neural Network and dynamically merges them with multimodal retrieval signals. To adapt to evolving dialogue, a Reflective Reranker module iteratively critiques and refines candidate tracks, outputting personalized recommendations and associated explanations. Experimental results show our pipeline outperforms baselines in ranking, diversity, and explanations, proving that combining graph embeddings with agentic reasoning successfully bridges long-term profiles and immediate intent.
