2026-09-09

1Deployed a generative recommendation model based on HSTU for product recommendation — Shopify

Here, a "generative" recommendation model means a sequence Transformer that treats a user's interaction sequence (search, view, add to cart, purchase) as a token sequence and predicts the next interaction autoregressively. It does not generate items as text; it predicts a product embedding from the catalog as the next token.

The base architecture is HSTU (Hierarchical Sequential Transduction Units), which Meta released in 2024. HSTU is a recsys-specific Transformer that places candidates and history in one sequence through an attention mask, so one forward pass can score several candidates at once. Shopify says it used this as a starting point and changed the details to fit the e-commerce context. This source is an attempt to add a generative recommender to the existing stack as both a candidate generator and one ranking candidate source.

• It is an autoregressive sequence predictor that takes the buyer's raw event sequence (search, view, add to cart, purchase) directly as input. Causal masking keeps it from referring to future events verified.
• For time information, it combines a rotary encoding inspired by RoPE (rotary positional embedding) with a relative attention bias, so attention can use both absolute time and relative recency verified.

"RoPE-inspired rotary encoding so attention can carry an absolute notion of time" — Shopify's Foundational Generative Recommender System, 2026-02-25

• To improve the quality of contrastive learning, it shares negatives (comparison examples that are not the correct answer) across batches, and introduced positive-aware hard negative selection, which picks hard negatives while knowing the positive verified.
• To raise recall step by step within the ensemble, it used a boosting-based approach that treats the predictions of other ensemble models as hard negatives verified.
  • This approach targets the problem that a new model picks only candidates that overlap with existing candidate generators and so adds nothing to the ensemble total; it puts candidate diversity directly into the training objective.
• CUDA kernel optimization raised pipeline training speed by up to 7.3x over the baseline verified.
  • This is an improvement in training throughput, not an architectural scaling law. We still read it as a sign that the serving and training cost of the HSTU family remains a tuning target.

"In our internal benchmarking, the full training pipeline reached up to 7.3x speedup" — Shopify's Foundational Generative Recommender System, 2026-02-25

• In the production A/B, Shop orders rose +0.94%, high-quality CTR +5%, conversion rate +0.71%, and final product recall@4 +4.8% (all relative) verified.
  • The baseline is the existing recommendation stack, and the specific model name was not disclosed.
• For serving scale, it gave only a metric from the platform-wide traffic context, not the number of product recommendation requests: 2.2 trillion edge requests handled during BFCM 2025 (Black Friday Cyber Monday) verified.
  • It describes the peak event scale as being in the billions.

"Shop orders increased by 0.94% relative. High quality click-through rate increased by 5% percent relative...Conversion rate increased by 0.71% relative." — Shopify's Foundational Generative Recommender System, 2026-02-25

• As a next step, it plans to introduce semantic IDs (SID) to represent a product as a token sequence drawn from a small token vocabulary, which would reduce dependence on large embedding tables and open up text query integration verified.
  • This direction has not yet been deployed.

"when products are modeled independently, the system relies on very large embedding tables...can become a bottleneck" — Shopify's Foundational Generative Recommender System, 2026-02-25

• This source was published in 2026-02, so it is not a fresh announcement, but it is a statement this repository had missed for over six months. It was found today through WebSearch cross-checking.
• Shopify is the sixth industry company attached to the HSTU lineage. The lineage is summarized in topics/sequence-transformer-ranking.md.
  • The existing lineage is Meta HSTU and Wukong originals, Netflix Foundation Model, Meta ads LLaTTE, Alibaba TransRetrieval, ByteDance ReST, and Tencent PCG CCFormer.
  • Until now, the statements were concentrated in three domains. In ads: Meta LLaTTE, Alibaba TransRetrieval, and ByteDance ReST. In streaming video: Netflix FM. In social and content: the Meta HSTU original and Tencent PCG.
  • Shopify is the first statement from a separate domain, e-commerce product recommendation.
  • It is also evidence that the HSTU license and the open-source reference implementation are accessible enough for other companies to adopt in production.
• In terms of SID adoption stage, Shopify sits at a midpoint. It deployed to production without SID, using large embedding tables, and plans to move to SID afterward.
  • Most recent statements in the generative retrieval / semantic ID thread are follow-up improvements on systems that already deployed SID: Kuaishou HD-Rec and PushDualGen, Alibaba PailitaoGR, Spotify GLIDE, and Tencent BARGE.
  • This is consistent with the observation that SID adoption proceeds in sequence at companies that actually feel the embedding-table scale problem in practice.
• The production A/B lift itself is not larger than in other statements (Shop orders +0.94%, conversion +0.71%).
  • However, it should be read together with the facts that e-commerce is a surface where absolute conversion figures are inherently low and that Shopify cites scale (2.2 trillion edge requests at BFCM) as its evidence.
  • Not disclosing what the baseline was counts against its reproduction value.
• There are also limits. Credit to the original HSTU is not stated explicitly, and the baseline model's identity and parameter scale are not public.
  • The text does not credit the original HSTU authors (Meta, Jiaqi Zhai et al., arXiv 2402.17152) or the company Meta, and only links an open-source tutorial.
  • Because the baseline model's identity and parameter scale are not public, the evidence it adds to the scaling law thread is weak. Whether these two pieces of information appear in follow-up sources (such as a RecSys industry track) is the point to watch.
    • The identity here means whether the baseline is a previous two-tower structure (one that embeds users and items separately), a DLRM family model, or a GBDT ranker.

Corrections

Nothing was overturned. This source added new evidence to the threads and does not conflict with existing conclusions.

Radar

• Added Shopify Generative Recommender to "New this month" on the radar (a watchlist of papers and trends outside the tracked services). It is a belated discovery of a 2026-02 announcement.
• Also added it to the "Recommendation model scaling laws" and "Generative retrieval / semantic ID" threads.

What to check next