Event Details
KG-ARMM: Residual Structural Fusion for Graph-Augmented Personalized Retrieval
Presenter: Mohamed Elrfaey
Supervisor:
Date: Thu, October 22, 2026
Time: 11:30:00 - 00:00:00
Place: Zoom - see below.
ABSTRACT
Join: https://uvic.zoom.us/j/84049527002?pwd=1bVYu8GGjTRBGl3Z8TG9MQouVdNIEP.1 | Meeting ID: 840 4952 7002 | Passcode: 872754
In case Uvic Zoom is still down, please use this link instead: https://us05web.zoom.us/j/81488037378?pwd=LhbTFk8nanQ3LpPNwT0bMPyTfIDHHC.1
Meeting ID: 814 8803 7378
Passcode: dz7i6Q
Title: KG-ARMM: Residual Structural Fusion for Graph-Augmented Personalized Retrieval
Abstract
Dense retrieval can find movies with similar text and imagery, but it can miss the relationships that connect a user’s interests: a shared director, a franchise, or a neighborhood of related items. Adding graph information is not automatically helpful either. A re-ranker can improve average accuracy while concentrating recommendations on already popular movies. This seminar presents KG-ARMM, a framework that adds structural evidence as a controlled correction to a fixed multimodal retriever. A text-and-poster backbone generates candidates, while an offline item graph and a filtered movie knowledge graph supply features to a lightweight pairwise re-ranker. The final score combines dense relevance, learned and structural support, and explicit popularity and long-tail terms. Stored entity paths make the structural evidence inspectable. On a temporal MovieLens–TMDB benchmark with 135,878 users and 8,608 movies, the balanced model improves NDCG@20 by 8.3% and Recall@20 by 7.0% over the multimodal backbone. Gains remain positive against a matched no-structure control, with essentially unchanged catalog coverage and long-tail exposure. Users with shorter histories benefit most. The talk will examine the accuracy–exposure tradeoff, the limits imposed by candidate recall, and the distinction between traceable evidence and validated explanations. Full-budget LightGCN remains the strongest absolute ranker in the reported benchmark; KG-ARMM targets a complementary problem: improving an existing retrieval pipeline without replacing it.
Contact elrefaey@uvic.ca