HD-Rec Pairs Hierarchical Quantization with Sparse Domain Experts for Cross-Domain Generative Recommendation, Up to 17.6% on Hit Rate

Hierarchical Quantization with Domain-Adaptive Sparse Routing for Generative Cross-Domain Recommendation

Haiying He, Xiaopeng Li, Yuchen Gu, Kuo Cai, Bo Chen, Jingtong Gao, Yejing Wang, Derong Xu, Ruiming Tang, Guorui Zhou, Han Li, Xiangyu Zhao

cs.IR

2026-08-07

HD-Rec pairs hierarchical quantization with sparse domain experts and routing consistency for cross-domain generative recommendation, leading three benchmarks with up to 17.6% on Hit Rate.

What problem this solves

Generative recommendation (GenRec) is a recent paradigm: compress each item into a few tokens called a Semantic ID (SID), then have a Transformer predict the next token to generate recommendations directly, skipping retrieval over a candidate catalog. It works well within a single domain. Extending it to cross-domain recommendation is harder, because item semantics and user behavior differ sharply across domains; apparel and electronics barely share a vocabulary.

Prior work either forces one globally shared representation across all domains, which lacks capacity, or applies lightweight domain adaptation that breaks down when the domain gap is large. HD-Rec's premise is that cross-domain semantics are hierarchical: coarse patterns (say "sports electronics" or "basketball") can be shared, while fine-grained preference differences need to be domain-specific. The whole framework is built around that premise.

Method

Three components cover representation, capacity, and consistency.

• Hierarchical Domain-aware Quantizer (HDQ): when building semantic IDs, the first L-1 layers use a globally shared codebook for coarse semantics, and the final layer switches to K domain-specific codebooks chosen by a lightweight Gumbel-Softmax router based on the item's residual. Shared coarse, domain-specific fine, matching the hierarchical premise. Ablations that swap this for all-shared codebooks, or put domain-specific codebooks first, drop H@10 on both domains.
• Domain-Adaptive Sparse MoE (DAS MoE): one always-on shared expert processes every token, plus K specialized experts; a gate selects one specialized expert per token via top-1 routing. Conditional capacity that varies by domain, at the cost of activating only one extra expert.
• Cross-Granularity Routing Consistency (CRCL): the L tokens of one item should share an expert, but standard MoE routes each token independently, which can send tokens of the same item to very different experts. CRCL pulls token-level routing distributions toward the item-level (mean-pooled) distribution via KL divergence.

Results

Three cross-domain datasets: Clothing-Sports (Leisure, Amazon), Electronics-Phones (Technology, Amazon), Books-Movies (Entertainment, Douban). Metrics: Hit Rate@10 and NDCG@10. HD-Rec is best on all 12 domain-metric pairs.

Gains concentrate on mid-to-low, high-gap domains (Sports, Electronics); Movies, already high in absolute terms, barely moves. Removing any of the three components hurts, with HDQ and DAS MoE contributing more than CRCL. Inference slows only about 2.0% to 2.5%, and routing variance drops about 6x, evidence that CRCL is doing its job.

Why it matters

Cross-domain recommendation has long been blocked by the need for overlapping users or items, which privacy and platform isolation make impractical. The generative plus hierarchical-representation route sidesteps that requirement, and HD-Rec shows that "shared coarse, domain-specific fine" is a cost-effective compromise. For recommendation engineers, this combination of hierarchical quantization and sparse experts is something to try directly on multi-business scenarios.

Limitations

The authors concede three points: experiments cover only two-domain settings, so stability and scalability with more or larger domains are unverified; the method depends on content-derived item representations and semantic-ID tokenization, which suffer when metadata is noisy or poorly aligned with real preferences; and only offline evaluation was done, with dynamic catalogs, interest drift, and shifting domain distributions left unaddressed. The near-zero gain on Movies, a dense high-absolute-value domain, also suggests diminishing returns where data is already abundant.

Terms

• Semantic ID (SID): a short encoding that compresses an item into a few discrete tokens; generative recommenders predict these tokens to produce recommendations.
• Codebook: a set of learnable vectors; quantization maps each input to its nearest vector, the basis of discretization.
• Residual quantization: quantize the error of the previous layer repeatedly, stacking layers for multi-granularity semantic IDs.
• MoE (Mixture-of-Experts): several sub-networks each handle a class of input, with a gate selecting dynamically; a standard way to grow capacity without linearly growing compute.
• Gumbel-Softmax: a differentiable continuous approximation of a discrete choice, letting a router train with gradients.
• next-token prediction: as in language models predicting the next word, here predicting the next semantic token of a recommended item.

Source

• arXiv 2608.06997v1
• PDF
• Hugging Face

What people are saying

• HD-Rec Tackles Cross-Domain Recommendation with Hierarchical Quantization & Sparse Routing _reachsumit

Related papers

• LARGER injects call-graph neighbors into grep and lifts LocBench Acc@5 from 74.1 to 87.0
• Meta hybrid GPU-CPU search: a billion-doc interaction path plus 20× CPU inventory
• Yonsei Self-Index lets a search index rewrite its own keys, +57% nDCG with BGE
• Alibaba CoFree stops reasoning collapse and lifts 4B embeddings by 2.8 nDCG@10
• Honor's MERIT-Rank: a 4B reranker beats most 7B and 32B models on BRIGHT
• Sparse Llama-3 retrievers beat dense ones; knowledge distillation stops scaling by 8B

All paper explainers