Frozen embeddings plus score fusion match semantic recommenders, beating them 7-12% on Amazon

Auditing Semantic Gains in Sequential Recommendation: A Lightweight Recovery Test

Kong Wang, Zhongke He, Xiang Chen, Hongwei Zeng, Kai Deng, Long Wang, Kehua Yang

cs.IR, cs.AI

2026-08-02

A transparent fusion of SASRec, ItemCF, and frozen BGE beats the strongest semantic-recommendation baselines by 7-12% on three Amazon datasets, with no serving-time language model.

What problem this solves

Recent semantic and generative recommenders routinely report large gains over ID-only sequential baselines. But these methods are evaluated as complete systems, so where the gain comes from is unclear: language-model reasoning, semantic-ID generation, end-to-end semantic architecture, stronger offline item representations, or simply that semantic and collaborative signals are complementary. This is not a pedantic attribution question. If most of the gain can be recovered with frozen offline embeddings and simple fusion, then stacking complex semantic machinery may not be necessary.

Method

LIME-Rec is not another recommendation architecture; it is an auditable recovery test. It fuses the scores of three independent experts: a sequential expert SASRec (interactions only), a collaborative expert ItemCF (co-occurrence over the most recent 20 interactions, time-decayed and sqrt-normalized to dampen popularity), and a semantic expert using frozen BGE (bge-base-en-v1.5) item embeddings, with a recency-weighted average of the user's history giving a cosine score. The three scores are min-max normalized per user, combined by a lightweight linear gate fitted only on validation data with softmax weights, and finished with a bounded history calibration that subtracts up to 0.1 from already-interacted items.

The defining property is that there is no serving-time language-model inference, no semantic-ID decoding, and no end-to-end text optimization anywhere; item text is encoded offline once. Every expert's contribution stays separately inspectable, and the fusion is transparent and score-level.

Results

Three text-rich Amazon datasets (Beauty, Toys, Sports), full-catalog evaluation with repeats allowed:

DatasetLIME-Rec R@10Strongest baseline (GRAM)Relative gain
Beauty0.09960.0890+12.0%
Toys0.11050.0987+12.0%
Sports0.05930.0554+7.0%

LIME-Rec tops all 12 dataset-metric combinations, beating TIGER, IDGenRec, HSTU, LC-Rec, and MHL. A factorial ablation shows three-expert fusion without calibration already consistently beats calibrated SASRec, so recovery is not explained by history calibration alone. The most telling result is the text-correspondence control: shuffling item-text embeddings across item IDs (preserving the count and dimensionality of the embeddings and the whole architecture) drops R@10 by 13.6% to 17.5% and lowers the gate's weight on the semantic expert in lockstep, proving the gain rests on genuine item-text correspondence rather than an extra score channel.

Why it matters

This is a measurement warning, the opposite of "build a stronger model." Its practical implication: before crediting recommendation gains to online language modeling, semantic-ID generation, or heavy semantic architecture, run a frozen-embedding-plus-transparent-fusion baseline and see how much it recovers. If it matches or beats the heavy system, those mechanisms are not empirically necessary on that benchmark. It also hands reviewers a ruler for judging how much metal is actually in new semantic-recommendation papers.

Limitations

The authors are careful to call this a mechanism-necessity test, not a causal decomposition: a lightweight system matching a heavy one does not mean the heavy model does the same thing internally, or that its mechanisms are useless elsewhere. Only three English, text-rich Amazon datasets are tested, with one frozen small encoder (BGE-base); cold-start, cross-domain, and sparse-behavior settings where language reasoning might genuinely help are not covered. The protocol is full-catalog with repeats allowed (maskhistory=false), a specific choice, and conclusions could shift under a different protocol. What it proves is that the benchmark score is attainable, not that heavy models are superfluous.

Terms

Source

What people are saying

Related papers

All paper explainers