URecJPQ : Memory-efficient Multimodal Recommendation Models through RecJPQ in Large-Scale Scenarios

Large-scale industrial recommendation systems face significant memory constraints due to the massive number of unique user and item ID embeddings required. While multimodal features improve recommendation quality, they further increase the memory footprint. This paper introduces URecJPQ, a method designed to compress these embeddings through Joint Product Quantization (JPQ) to enable efficient training in large-scale, multimodal top-k recommendation scenarios.

URecJPQ extends the existing RecJPQ framework—originally developed for sequential recommendation—to the traditional top-k setting. Instead of learning unique, high-dimensional embeddings for every user and item, URecJPQ learns a set of shared, low-dimensional sub-embeddings. Each entity is represented by a concatenation of these sub-embeddings, which are mapped via discrete codes stored in a codebook. The authors propose two strategies for code assignment: a random strategy and a Truncated SVD-based strategy that preserves semantic consistency by clustering users and items based on their interaction history. Because the quantization occurs before training, the backbone model can be trained end-to-end.

Experiments across three datasets (Movies, Baby products, and Sports products) demonstrate that URecJPQ drastically reduces model size. The authors report a 98% to 99% reduction in the number of trainable parameters and an 86% to 98% reduction in checkpoint sizes. This massive compression comes with only marginal decreases in accuracy (averaging 8.5% in Recall and 16% in NDCG), and in some specific domains like baby products, the model even achieved performance improvements of up to 85% compared to uncompressed baselines.