PaperlayerView exact revision on arXiv

Source-linked AI summary

Rethinking Generative Recommender Tokenizer: Recsys-Native Encoding and Semantic Quantization Beyond LLMs

Yu Liang, Zhongjin Zhang, Yuxuan Zhu, Kerui Zhang, Zhiluohan Guo, Wenhang Zhou, Zonqi Yang, Kangle Wu, Yabo Ni, Anxiang Zeng, Cong Fu, Jianxin Wang, Jiazhi Xia

arXiv:2602.02338v1cs.IRcs.AI

TL;DR

Existing SID pipelines rely on semantic encoders and generic quantization, creating a mismatch with recommendation objectives and weakening sequential predictability. ReSID addresses this with recommendation-native representation learning and objective-aligned quantization, outperforming strong baselines while reducing tokenization costs.

  • Problem

    Existing SID pipelines use foundation-model embeddings and generic quantization, but their semantic objectives are weakly coupled with recommendation and their quantization can weaken sequential predictability.

  • Method

    ReSID aligns representation learning and quantization with generative recommendation by using mutual-information-based task-aware representations and jointly minimizing reconstruction and prefix-conditional entropy.

  • Results

    ReSID consistently outperforms strong sequential recommenders and SID-based generative models by over 10% while reducing tokenization costs by up to 122×.

  • Takeaways & Limitations

    ReSID produces compact and predictable SIDs efficiently without foundation models and demonstrates superiority through theoretical analysis and extensive empirical results.

  • Takeaways & Limitations

    SID-based generative models converge substantially more slowly than item-ID-based methods such as SASRec, and principled diagnostics for GAOQ remain open.

Abstract

from arXiv · show

Semantic ID (SID)-based recommendation is a promising paradigm for scaling sequential recommender systems, but existing methods largely follow a semantic-centric pipeline: item embeddings are learned from foundation models and discretized using generic quantization schemes. This design is misaligned with generative recommendation objectives: semantic embeddings are weakly coupled with collaborative prediction, and generic quantization is inefficient at reducing sequential uncertainty for autoregressive modeling. To address these, we propose ReSID, a recommendation-native, principled SID framework that rethinks representation learning and quantization from the perspective of information preservation and sequential predictability, without relying on LLMs. ReSID consists of two components: (i) Field-Aware Masked Auto-Encoding (FAMAE), which learns predictive-sufficient item representations from structured features, and (ii) Globally Aligned Orthogonal Quantization (GAOQ), which produces compact and predictable SID sequences by jointly reducing semantic ambiguity and prefix-conditional uncertainty. Theoretical analysis and extensive experiments across ten datasets show the effectiveness of ReSID. ReSID consistently outperforms strong sequential and SID-based generative baselines by an average of over 10%, while reducing tokenization cost by up to 122x. Code is available at https://github.com/FuCongResearchSquad/ReSID.

1. Introduction

Existing SID pipelines use semantically oriented embeddings and generic quantization, creating misalignment with collaborative recommendation and sequential decoding. ReSID addresses these issues through recommendation-native representation learning and objective-aligned quantization.

  • Motivation: Semantic-centric SID pipelines embed items with foundation models before generic discretization, enabling compact token sequences but misaligning tokenization with recommender objectives.These pipelines replace atomic item IDs with short discrete sequences for autoregressive prediction.
  • Motivation: Foundation-model embeddings emphasize semantic similarity, which can conflict with collaborative signals and produce representations that are neither semantically clean nor optimally aligned with recommendation.Frequently co-occurring items may differ substantially in semantic or visual properties.
  • Motivation: Existing quantizers often assign child indices locally and independently, causing semantic ambiguity and weaker reconstruction fidelity across hierarchical SID codes.Different item groups can share the same token at a level despite differing semantics.
  • Contributions: ReSID introduces FAMAE, which learns recommendation-sufficient field-level representations through masked structured-feature prediction.Its information-theoretic formulation targets information relevant to downstream recommendation.
  • Contributions: GAOQ jointly reduces reconstruction error and prefix-conditional uncertainty through globally consistent indexing across hierarchical levels.The resulting compact codes are designed to improve sequential predictability.
  • Results: Across ten public datasets, ReSID achieves over 10% relative improvement while reducing tokenization costs by up to 122× on million-scale datasets.It outperforms strong sequential recommenders and state-of-the-art SID-based generative models.

2. Rethinking SID Pipeline

SID recommendation comprises encoding, quantization, and autoregressive generation, so information loss and target-code quality connect upstream design to downstream prediction. ReSID therefore emphasizes collaboration-dominant representations and predictable, information-preserving codes.

  • Three-Stage SID Pipeline: A SID recommender maps item features to continuous representations, discretizes them into finite-length code sequences, and predicts target codes from user history.The three components are an encoder, quantizer, and generative model.
  • Three-Stage SID Pipeline: Because quantized SIDs become autoregressive training targets, noisy or inconsistent codes distort the supervision signal without downstream correction.The encoder–quantizer pair therefore determines the quality and semantic consistency of the generated targets.
  • Design Criteria: An effective E-stage should prioritize collaborative interaction signals while retaining semantic information as auxiliary context, because discretization can blur or discard task-relevant information.The paper proposes evaluating representation quality with task-aware metrics.
  • Design Criteria: The Q-stage should preserve task-relevant information while reducing intrinsic uncertainty in autoregressive decoding through sequentially predictable and semantically stable codes.Reconstruction distortion alone is insufficient for SID quantization.
  • ReSID: ReSID aligns both stages with generative recommendation by combining collaboration-dominant representations with compact, autoregressive-decoding-friendly SIDs.The framework uses FAMAE for representation learning and GAOQ for quantization.

3. Methodology

ReSID learns field-aware representations from structured features and quantizes them with globally aligned codes. Its objectives and diagnostics target predictive sufficiency, semantic structure, information preservation, and sequential predictability.

  • Framework: ReSID combines FAMAE for collaboration-dominant item representations with GAOQ for compact, autoregressive-decoding-friendly Semantic IDs.The framework is designed from an information-theoretic perspective.
  • Field-Aware Masked Auto-Encoding: FAMAE masks target-item fields and predicts them from remaining fields and user history using a field-aware Transformer objective.Field-specific prediction preserves separable supervision across structured features.
  • Field-Aware Masked Auto-Encoding: The FAMAE loss increases a variational lower bound on mutual information between the contextual representation and target-item features.This provides an information-theoretic proxy for predictive sufficiency.
  • Item Representation Extraction: Field-level embeddings, rather than history-entangled contextual representations, form the basis for SID quantization because they retain item-feature information while remaining independent of user-specific context.The paper uses concatenated field embeddings as input to quantization.
  • Task-Aware Metrics: FAMAE evaluates embeddings with complementary proxies for collaborative predictability and discriminative semantic structure.These correspond to full-field masking for target prediction and single-field masking for item-ID prediction.
  • Task-Aware Metrics: Representations performing well on both metrics consistently yield higher-quality Semantic IDs and improved downstream generative recommendation performance.The relationship is reported empirically in Figure 3.
  • Globally Aligned Orthogonal Quantization: GAOQ jointly optimizes quantization objectives through hierarchical vector quantization and globally aligned indexing, producing compact, semantically stable, autoregression-friendly codes.Its design combines balanced clustering with globally shared reference directions.

4. Experimental Settings

The experiments evaluate ReSID on ten Amazon-2023 review subsets against item-ID, feature-augmented sequential, and SID-based generative baselines. Recommendation uses Recall@K and NDCG@K under leave-one-out evaluation.

  • Research Questions: The experiments examine overall performance, component contributions, task-aware metric validity, and SID tokenization efficiency.These correspond to four stated research questions.
  • Datasets: ReSID is evaluated on ten subsets of the Amazon-2023 review dataset spanning categories such as Musical Instruments, Video Games, and Books.The study follows standard preprocessing practice and reports dataset details in Appendix C.
  • Compared Methods: The comparison addresses a confounder in prior SID evaluations, where SID systems use rich metadata while sequential baselines use item IDs only.The authors include feature-augmented sequential models for a fairer comparison.
  • Compared Methods: The study compares item-ID-only sequential recommenders, structured-feature sequential variants, and SID-based generative recommenders.Baselines include HGN, SASRec, BERT4Rec, S3-Rec, TIGER, LETTER, and EAGER.
  • Evaluation Settings: Recommendation performance is measured with Recall@K and NDCG@K for K ∈{5, 10} using leave-one-out testing and validation.The final interaction is used for testing and the penultimate interaction for validation.

5. Experimental Results

Across ten Amazon-2023 datasets, ReSID consistently outperforms sequential and SID-based baselines, including fair comparisons with structured side information. Ablations support the complementary roles of FAMAE and GAOQ, while task-aware metrics track downstream performance and quantization is substantially faster.

  • Overall performance: ReSID consistently achieves the best performance across all metrics and datasets, outperforming item-ID sequential recommenders even with structured side information.
  • Impact of structured features: Structured-feature augmentation substantially improves sequential recommenders, with SASRec* and BERT4Rec* often matching or exceeding prior SID pipelines.The comparison indicates that some earlier SID gains may reflect additional side information rather than tokenization alone.
  • ReSID vs. prior SID-based methods: 16.0%/13.8% average improvements over LETTER on Recall@5/10 and 16.2%/14.9% on NDCG@5/10 establish ReSID’s advantage over prior SID methods.
  • Ablation study: ReSID outperforms semantic, collaborative, and standard sequence-encoder replacements, supporting FAMAE’s predictive sufficiency and preservation of structured feature identity.
  • Ablation study: ReSID outperforms RQ-VAE and Hierarchical K-Means variants, indicating that GAOQ improves sequential predictability and benefits from global index alignment beyond reconstruction error alone.
  • Task-aware embedding metrics: Downstream R@10 increases as both task-aware embedding metrics improve across Musical Instruments and Baby Products, without repeated end-to-end retraining.
  • Efficiency of SID tokenization: 77×–122× more runtime is required by LETTER than ReSID for quantization, while TIGER incurs approximately a 5× overhead relative to ReSID.
  • Efficiency of SID tokenization: Runtime comparisons omit representation-learning time, limiting the efficiency comparison to the quantization stage.FAMAE is described as comparable in cost to light sequential models, whereas prior encoder costs may be amortized or unspecified.

6. Limitations and Future Work

The authors identify unresolved diagnostic and convergence challenges for SID-based recommendation. They leave these issues for future work.

  • Principled diagnostics for GAOQ remain an open challenge despite FAMAE providing task-aware embedding-quality metrics.
  • SID-based generative models converge tens of times more slowly than item-ID-based methods such as SASRec.

7. Conclusion

ReSID is a recommendation-native SID framework that aligns representation learning and quantization with generative recommendation’s information requirements. FAMAE and GAOQ produce compact, predictable SIDs efficiently without foundation models.

  • ReSID aligns representation learning and quantization objectives with the information requirements of generative recommendation.
  • FAMAE and GAOQ produce compact and predictable SIDs efficiently without foundation models.
  • Theoretical analysis and extensive empirical results demonstrate ReSID’s superiority across the reported evaluations.
  • ReSID is presented as the first SID-based approach to outperform strong item-ID baselines augmented with side information.

Impact Statement

The paper advances methodological work in recommender-system representation learning and discrete tokenization. Its techniques are intended to improve recommendation performance and scalability, with no significant direct societal or ethical consequences anticipated.

  • The paper advances machine learning for recommender systems through representation learning and discrete tokenization techniques.
  • The proposed methods are intended to improve the performance and scalability of recommendation models.
  • The authors do not foresee significant negative societal or ethical consequences arising directly from this work.

C. Dataset

The experiments use a standard 5-core sequential-recommendation setting with chronologically ordered interactions and leave-one-out evaluation. Training uses sliding windows with maximum sequence length 32.

  • The dataset processing removes users and items with fewer than five interactions under the standard 5-core setting.
  • Remaining interactions are chronologically ordered into user behavior sequences and evaluated using leave-one-out.
  • Training uses a sliding-window strategy with a maximum sequence length of 32.
  • Table 6 is identified as reporting dataset statistics.

D. Compared Methods

The evaluation compares ReSID with sequential recommendation models and SID-based generative methods using standard benchmark settings and reported branching-factor configurations.

  • ReSID is compared with representative sequential recommendation models and recent generative recommendation methods.
  • Sequential recommendation methods: HGN captures short- and long-term interests through hierarchical gating over interaction sequences.
  • Sequential recommendation methods: SASRec uses unidirectional self-attention to model sequential dependencies and predict the next item.
  • Sequential recommendation methods: BERT4Rec learns sequence representations through bidirectional masked item prediction using left and right context.
  • Generative recommendation methods: TIGER, LETTER, EAGER, UNGER, and ETEGRec represent generative recommendation alternatives using semantic identifiers, collaborative signals, or unified tokenization and recommendation objectives.
  • Table 8 varies first- and second-level branching factors and reports Recall@10, with each row fixing b1 while sweeping b2.

E. Implementation Details

The experiments use a three-stage pipeline consisting of representation learning, SID quantization, and SID-based generative modeling, with shared controlled configurations for comparison.

  • The experimental pipeline comprises representation learning (E-stage), SID construction via quantization (Q-stage), and SID-based generative modeling (G-stage).
  • E-stage and sequential baselines: E-stage encoders and sequential baselines use matched Transformer configurations, including embedding/hidden size 128, two layers, four attention heads, and FFN dimension 512.
  • Q-stage: ReSID uses dataset-specific branching factors, while SID baselines follow optimal quantization settings from their original papers and official implementations.
  • G-stage: All SID-based generative recommenders use the same T5-style encoder-decoder architecture with four encoder and four decoder layers.

F. Full Results of Main Experiments

The full experiments examine performance across Amazon-2023 subsets, branching-factor sensitivity, scaling behavior, and the structure and predictability of learned representations and codes.

  • Table 4 reports absolute performance for all compared methods across Amazon-2023 subsets, complementing macro-averaged relative improvements.
  • Branching-factor sensitivity: At moderate branching factors, (b1, b2) = (32, 40) achieves the highest Recall@10 in the Musical Instruments sweep.
  • Model scaling: ReSID consistently achieves the best NDCG@10 across the explored parameter range, although the final largest-scale point slightly drops.
  • Semantic–collaborative alignment: FAMAE is the only compared approach producing well-structured clusters under both semantic category and behavioral community views.
  • Embedding structure: FAMAE aligns item-ID embeddings with category structure, whereas BERT4Rec item-ID embeddings are largely unstructured.
  • Embedding structure: FAMAE jointly supervises multiple fields to produce a structured, field-aligned embedding space for downstream SID construction.
  • Contextual aggregation: FAMAE attention shows recency bias toward recent interactions, matching the information requirements of next-item prediction.
  • SID predictability: FAMAE representations yield higher SID overlap ratios, while GAOQ’s global index alignment further increases overlap and reduces prefix-dependent ambiguity.
Loading 2602.02338v1…