Please note: This PhD seminar will take place in DC 3301.
Liam Hebert, PhD candidate
David R. Cheriton School of Computer Science
Supervisors: Professors Robin Cohen & Lukasz Golab
A prominent trend in recommender systems involves replacing ID-based collaborative filtering models with architectures that rely solely on textual item descriptions. This work begins by challenging the premise of this trend, demonstrating that a thoroughly tuned ID-only baseline (Bert4Rec) is significantly more competitive than previously reported. This finding establishes the continued importance of collaborative signals and motivates our proposal of a hybrid solution.
We introduce FLARE (Fusing Language models and collaborative Architectures for Recommender Enhancement), a novel hybrid recommender that integrates a collaborative filtering model with a language model using a Perceiver network. To ensure the practical relevance of our findings, we evaluate FLARE on both standard and large-scale, high-vocabulary datasets, establishing new benchmarks for the latter. Furthermore, we propose and evaluate critiquing as a novel task to assess a model's ability to incorporate explicit user intent, a critical form of contextual understanding. Our results show that FLARE achieves strong performance and effectively integrates user intent, validating the superiority of a hybrid approach that unifies both collaborative and semantic signals.