Higher Capacity Transfer into JPQ Item Codes for Scalable Sequential Recommendation
Year: 2026
Type: conference-paper
Abstract: Although transformer models have improved the effectiveness of sequential recommendation, their computational and memory costs make it challenging to scale to item catalogues containing millions of items. To address this limitation, RecJPQ, a recommendation method based on joint product quantisation, represents each item using compact discrete codes rather than large embedding tables, substantially reducing training memory requirements. These codes are typically initialised from a memory-efficient bootstrap representation, such as truncated SVD, before being quantised. However, this process introduces an information bottleneck: because quantisation is inherently lossy, information may be lost when the bootstrap representation is mapped to the fixed-budget discrete item-code space, resulting in suboptimal code assignments. To address this capacity transfer problem, we initialise JPQ item codes using a higher-dimensional bootstrap representation and adopt product quantisation (PQ) to compress this representation into the fixed-budget discrete item-code space prior to training; we refer to this method as RecJPQ-HC (Higher Capacity). Experimental results show that RecJPQ-HC consistently improves downstream recommendation performance over the original RecJPQ code assignment strategy. Across multiple datasets and sequential recommendation models, it achieves statistically significant average relative improvements of approximately 30% in NDCG@10 while preserving memory efficiency.
Institution: University of Glasgow
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203511005