Search papers, labs, and topics across Lattice.
This paper introduces LIME-Rec, a lightweight auditing framework designed to dissect the contributions of various recommendation strategies in sequential recommendation systems. By employing a combination of sequential, co-occurrence, and semantic experts, LIME-Rec reveals that significant performance gains in semantic recommenders are not solely due to advanced language modeling but rather depend on the genuine correspondence between item representations and their textual descriptions. The findings demonstrate that the proposed method outperforms traditional baselines while providing a transparent mechanism for evaluating the sources of improvement, emphasizing the importance of understanding the underlying mechanisms in recommendation systems.
Gains in semantic recommendation systems hinge more on item-text correspondence than on complex language modeling, challenging assumptions about model superiority.
Recent semantic and generative-retrieval recommenders report substantial improvements over ID-only sequential baselines, but it remains unclear whether these gains arise from language-model reasoning, semantic-ID generation, end-to-end semantic architectures, stronger offline item representations, or complementary semantic and collaborative signals. We investigate this attribution ambiguity through LIME-Rec, a lightweight and auditable recovery test. LIME-Rec combines three independent experts: a SASRec sequential expert, an ItemCF co-occurrence expert, and a semantic expert based on frozen BAAI/bge-base-en-v1.5 item embeddings. Their full-catalog scores are normalized per user and combined through auditable score-level fusion followed by bounded history calibration. The fusion gate and calibration head are fitted on validation data only, require no serving-time language-model inference, and keep each expert contribution separately inspectable. On Amazon Beauty, Toys, and Sports, LIME-Rec achieves R@10 scores of 0.0996, 0.1105, and 0.0593, outperforming the strongest comparison baseline by 7.0%-12.0%. Three-expert fusion without history calibration consistently outperforms calibrated SASRec, showing that calibration alone does not explain the recovery. Randomly permuting item-text embeddings across item IDs reduces R@10 by 13.6%-17.5%, indicating that the gains depend on genuine item-text correspondence rather than additional representation capacity. These results suggest that lightweight recovery from offline item representations and transparent fusion should be ruled out before improvements are attributed to serving-time language modeling, semantic-ID generation, or heavier semantic machinery.
