Pith. sign in

REVIEW 6 cited by

On the Embedding Collapse when Scaling up Recommendation Models

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2310.04400 v2 pith:5XAWQHPF submitted 2023-10-06 cs.LG cs.IR

classification cs.LGcs.IR
keywords modelscollapseembeddingrecommendationscalabilityinteractionanalysisavailable
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Recent advances in foundation models have led to a promising trend of developing large recommendation models to leverage vast amounts of available data. Still, mainstream models remain embarrassingly small in size and na\"ive enlarging does not lead to sufficient performance gain, suggesting a deficiency in the model scalability. In this paper, we identify the embedding collapse phenomenon as the inhibition of scalability, wherein the embedding matrix tends to occupy a low-dimensional subspace. Through empirical and theoretical analysis, we demonstrate a \emph{two-sided effect} of feature interaction specific to recommendation models. On the one hand, interacting with collapsed embeddings restricts embedding learning and exacerbates the collapse issue. On the other hand, interaction is crucial in mitigating the fitting of spurious features as a scalability guarantee. Based on our analysis, we propose a simple yet effective multi-embedding design incorporating embedding-set-specific interaction modules to learn embedding sets with large diversity and thus reduce collapse. Extensive experiments demonstrate that this proposed design provides consistent scalability and effective collapse mitigation for various recommendation models. Code is available at this repository: https://github.com/thuml/Multi-Embedding.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 6 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. ROCS: Request-Oriented Compute Sharing for Efficient Large-Scale Recommendation

    cs.LG 2026-07 conditional novelty 7.0 of 10

    ROCS restructures recommendation models so user-side computation is shared across all candidate items, yielding up to 3x serving throughput at equal or better prediction quality.

  2. Scaling Transformers for Discriminative Recommendation via Generative Pretraining

    cs.IR 2025-06 conditional novelty 7.0 of 10

    Generative pretraining plus sparse-embedding freezing makes large Transformer ranking models scale consistently, following a power law from 13K to 0.3B dense parameters.

  3. SpecFormer: Mitigating Embedding and Attention Collapse via Spectral-Aware Transformer for Recommendation

    cs.IR 2026-07 conditional novelty 6.0 of 10

    SpecFormer is a spectral-aware Transformer that flattens the singular-value spectrum of embeddings to prevent embedding/attention collapse, outperforming baselines on CTR benchmarks and scaling with layer depth.

  4. FuXi-\beta: Towards a Lightweight and Fast Large-Scale Generative Recommendation Model

    cs.IR 2025-08 conditional novelty 6.0 of 10

    FuXi-β shows that removing query-key attention and using a functional relative time bias makes generative recommendation Transformers faster and, on industrial datasets, more accurate.

  5. MTGR: Industrial-Scale Generative Recommendation Framework in Meituan

    cs.IR 2025-05 conditional novelty 6.0 of 10

    MTGR augments an HSTU-style generative ranking model with DLRM cross features and user-level aggregation, and reports a successful industrial deployment at Meituan with offline and online gains.

  6. Large Foundation Model for Ads Recommendation

    cs.LG 2025-08 conditional novelty 5.0 of 10

    Tencent's LFM4Ads transfers user, item, and user-item cross representations from a pre-trained foundation model into downstream ad models via feature, module, and model-level mechanisms, reporting a 2.45% platform-wid...

Pith tools