PaperlayerView exact revision on arXiv

Source-linked AI summary

FineSID: Scalable and Efficient Semantic Identifier Learning for Generative Recommendation

Song-Li Wu, Weinan Gan, Zhaocheng Du, Xianquan Wang, Jingyi Wang

arXiv:2609.36670v2cs.AI

TL;DR

Generative recommendation needs Semantic Identifiers that are scalable to large catalogs and efficient to generate, but Top-1 quantization leaves many codewords under-trained. FineSID spreads learning across the codebook while preserving semantic consistency and discrete generation. Across multiple public benchmarks, it improves codebook utilization and recommendation accuracy, particularly under long-tailed item distributions.

  • Problem

    Top-1 hard assignment concentrates gradients on frequently selected codewords, while existing SID solutions can trade semantic fidelity or inference efficiency against collision reduction.

    Source: Introduction · pp. 1–2

  • Method

    FineSID combines global, fine-grained codebook learning with discrete generation, using GLQ and a Semantic Consistency Module to support balanced optimization and semantic alignment.

    Source: Introduction · p. 2

  • Results

    FineSID consistently outperforms existing SID-based methods in codebook utilization and recommendation accuracy across multiple public benchmarks, with pronounced gains under long-tailed item distributions.

    Source: Introduction · p. 2

  • Takeaways & Limitations

    FineSID offers a principled SID-learning approach that jointly targets scalability and efficiency while preserving semantic structure.

    Source: Introduction · p. 2

Abstract

from arXiv · show

A critical prerequisite of generative recommendation is designing semantic identifiers (SIDs) that are both scalable to large item sets and efficiently learnable. Existing SID learning methods fundamentally rely on Top-1 hard assignment during vector quantization. While heuristic strategies -- such as clustering-based initialization or forced post-hoc collision resolution -- can artificially inflate codebook coverage, they often disrupt end-to-end semantic alignment and fail to address the underlying optimization bottleneck: sparse gradient propagation. In standard Top-1 assignment, gradients concentrate on a narrow subset of frequently selected codewords, leaving the majority inherently under-trained and causing severe SID collisions. To overcome this limitation natively without relying on complex initialization priors, we propose FineSID, a unified quantization framework that moves beyond Top-1 assignment by enabling fine-grained gradient propagation across the entire codebook. Instead of updating only a single selected codeword, FineSID distributes learning signals to all codewords in a soft, differentiable manner. This design promotes globally balanced codebook optimization while strictly preserving semantic consistency, effectively alleviating SID collisions and stabilizing training in large, high-dimensional codebooks. Extensive experiments on multiple public benchmarks demonstrate that FineSID is robust to initialization configurations and consistently improves both codebook utilization and recommendation accuracy. Our work provides a principled, initialization-agnostic solution for semantic identifier learning, advancing the practicality of generative recommendation.

Introduction

Generative recommendation uses Semantic Identifiers (SIDs) to turn items into discrete tokens, but Top-1 assignment concentrates gradients on a few codewords. FineSID instead optimizes across the codebook while preserving semantic alignment and discrete generation, targeting scalable and efficient SIDs.

Introduction

  • SIDs bridge continuous item representations and discrete token generation, providing a foundation for generative recommendation.Autoregressive recommendation frames interactions as sequence generation, with SIDs learned through quantization codebooks such as RQ-VAE (Wang et al. 2024).Source: Introduction · p. 1
  • Top-1 assignment concentrates gradients on frequently selected codewords, leaving most under-trained; initialization and forced reassignment do not resolve this bottleneck.Reassignment to semantically distant codes can disrupt semantic fidelity, while static initialization cannot prevent collapse during end-to-end training.Source: Introduction · pp. 1–2
  • Existing SID approaches can trade scalability against efficiency: regularization may suppress semantic differentiation, while auxiliary tokens lengthen decoding sequences.These limitations motivate globally optimizing the code space while preserving semantic structure.Source: Introduction · p. 2
  • FineSID distributes learning signals across the codebook while preserving discrete identifiers for autoregressive generation.Its Global–Local Quantization (GLQ) and Semantic Consistency Module support stable optimization and semantic alignment, particularly for long-tailed items.Source: Introduction · p. 2
  • FineSID consistently outperforms existing SID-based methods in codebook utilization and recommendation accuracy across multiple public benchmarks, with pronounced gains under long-tailed item distributions.The paper presents this as evidence of robust performance across the evaluated benchmarks.Source: Introduction · p. 2

Related Work

Generative recommendation represents items with discrete Semantic IDs and generates recommendations autoregressively. SID learning must balance uniqueness, semantic fidelity, and efficient generation, but existing approaches typically meet only some of these constraints.

Generative Recommendation Systems

  • Generative recommendation represents items with discrete Semantic IDs (SIDs) in an autoregressive sequence-generation framework.SID quantization maps continuous representations to discrete generation tokens and affects uniqueness, semantic fidelity, and recommendation performance.Source: Generative Recommendation Systems · p. 2

SID Collision in Generative Recommendation

  • SID collisions occur when distinct items map to identical SID sequences, creating ambiguous item grounding.Existing approaches use training-time regularization to balance codeword usage or auxiliary tokens to expand the representational space, but these address only parts of SID learning’s dual constraints.Source: SID Collision in Generative… · p. 2

Methodology

FineSID combines global–local quantization with semantic consistency to optimize SIDs while maintaining discrete generation and semantic fidelity. Its training formulation predicts the next item as a token sequence, using complementary mechanisms and objectives for codebook utilization, alignment, and ranking.

Methodology

  • FineSID combines Global–Local Quantization (GLQ) and the Quantization Semantic Consistency Module (QSCM) for scalable, efficient, semantically faithful SID learning.GLQ balances usage through local dense updates and global usage tracking; QSCM keeps continuous embeddings aligned with discrete SIDs.Source: Methodology · p. 2

Problem Formulation and Background

  • The task is to predict the next item from an interaction-history sequence, formulated as token-based sequence generation.Items are represented using content-derived embeddings and tokenized into Semantic Identifiers (SIDs) with a residual quantization tokenizer.Source: Problem Formulation and Background · p. 3

Global-Local Quantization (GLQ)

  • GLQ addresses sparse Top-1 updates with Local Refinement Quantization (LRQ) for dense gradient flow and Global Anchor Quantization (GAQ) for frequency balancing.LRQ uses soft assignment probabilities during backpropagation while the straight-through estimator preserves hard discrete assignments; GAQ adjusts update rates using historical codeword usage.Source: Global-Local Quantization (GLQ) · pp. 3–4

Quantization Semantic Consistency Module (QSCM)

  • QSCM enforces bidirectional alignment between continuous embeddings and discrete semantic anchors to limit semantic drift as GLQ updates the codebook.Its loss pulls quantized semantic anchors toward the original embedding and anchors the continuous embedding to its quantized counterpart.Source: Quantization Semantic Consistency… · p. 4

Optimization Objectives

  • FineSID jointly optimizes autoregressive ranking, item-representation reconstruction, and quantization semantic consistency objectives.The ranking loss aligns SIDs with sequential interactions, reconstruction retains item content, and QSCM aligns the continuous and quantized spaces.Source: Optimization Objectives · p. 4

Experiment Setup

Experiments use three public review datasets with a five-core filter and chronological user sequences. FineSID is compared with classical and generative recommendation baselines under leave-one-out, full-ranking evaluation.

Dataset

  • The evaluation uses Musical Instruments, Video Games, and Industrial Scientific reviews, filtered to users and items with at least five interactions and sequences capped at 50 items.Interactions span May 1996 through September 2023; sequences are ordered chronologically.Source: Experiment Setup · p. 4

Baseline Models

  • Generative baselines cover tokenizers trained with auxiliary objectives and tokenizers that expand the item vocabulary with additional tokens.Classical recommendation models are also included.Source: Experiment Setup · p. 5

Evaluation Settings

  • Evaluation uses Recall and NDCG at K=5 and 10, with each user’s latest interaction held out for testing and full-item ranking.The second-latest interaction is used for validation; generative models use beam search size 20.Source: Experiment Setup · p. 5

Implementation Details

  • Generative models use T5 with 6 encoder and 6 decoder layers; the RQ-VAE tokenizer has 3 codebooks of 256 embeddings each.All models use 128-dimensional embeddings; optimization uses AdamW with weight decay 0.05.Source: Experiment Setup · p. 5

Overall Performance

FineSID achieves the best recommendation performance across all three public benchmarks and all reported evaluation metrics. The authors attribute its gains to global optimization of the identifier space, which supports more balanced and semantically coherent tokenization.

Overall Performance

  • FineSID achieves the best performance across all three benchmarks and every evaluation metric, with statistically significant improvements over baselines.The reported comparisons are summarized in Table 2.Source: Overall Performance · p. 5
  • Among traditional sequential models, FDSA performs strongest, while methods adding identifier tokens generally outperform methods using auxiliary training objectives.The paper attributes FDSA’s result to additional textual embeddings and argues auxiliary objectives can distort semantic information.Source: Overall Performance · p. 5
  • Adding identifier tokens improves accuracy but increases identifier-space size, computational overhead, and reduces efficiency.Source: Overall Performance · p. 5

Ablation Study

The ablation study finds that the complete FineSID model outperforms its variants, with GLQ contributing the largest impact. The other modules provide complementary gains tied to semantic consistency, fine-grained distinctions, and SID validity.

Ablation Study

  • Removing GLQ causes the largest performance drop among module ablations, identifying it as the principal global optimization mechanism for SID learning.GLQ enables more effective aggregation of collaborative signals than local Top-1 assignment.Source: Ablation Study · p. 6
  • The full model consistently outperforms all ablation variants across datasets, supporting the combined design.Source: Ablation Study · p. 6
  • QSCM maintains semantic consistency, LRQ captures fine-grained semantic distinctions, and GAQ supports the semantic validity of each SID.The modules provide complementary gains.Source: Ablation Study · p. 6

Further Analysis

Further analyses report balanced codebook use, robustness to initialization, faster convergence, and stable utilization at larger codebook sizes. FineSID also outperforms selected baselines for seen and unseen users, while a toy trajectory illustrates broader codebook updates.

Codebook Analysis

  • FineSID achieves a balance rate of 0.4812 without auxiliary objectives or additional IDs, alongside continuous non-zero activation across the codebook.The paper links this balance to reduced redundancy and richer semantic variation.Source: Codebook Analysis · p. 6

Impact of Initialization on Semantic ID Generation

  • Random initialization performs competitively and often beats structured schemes; FineSID remains robust to initialization through progressive global codebook refinement.Among non-random strategies, collaborative initialization consistently outperforms semantic initialization.Source: Impact of Initialization on Semantic… · p. 6, Table 4 · p. 6

Efficiency Evaluation

  • FineSID reduces training time by more than 50% versus CAR across all three datasets while maintaining or improving recommendation performance.The comparison uses identical optimizer, learning-rate schedule, and early-stopping conditions.Source: Efficiency Evaluation · p. 6, Table 5 · p. 7

Scalability under High-Dimensional Settings

Generalizability Evaluation.

  • FineSID outperforms LETTER, SaviorRec, and CAR for both seen and unseen users in the evaluated Instrument and Scientific datasets.The unseen group comprises the 5% of users with the least interaction history.Source: Generalizability Evaluation · p. 7, Figure 4 · p. 7

Optimization Trajectory Analysis.

  • A toy optimization trajectory contrasts localized Top-1 updates with FineSID’s coordinated updates across identifiers for head and tail items.FineSID’s GLQ and QSCM jointly move identifiers toward task-aligned representations.Source: Optimization Trajectory Analysis · p. 7, Figure 5 · p. 7

Conclusions

FineSID addresses Top-1 assignment’s optimization bottleneck with global codebook gradient propagation while preserving discrete generation. The authors report balanced code usage and semantic fidelity, with consistent gains particularly for long-tailed and large-scale scenarios.

  • Top-1 hard assignment is identified as a SID-learning bottleneck that heuristic workarounds fail to resolve without semantic distortion.Source: Conclusions · pp. 7–8
  • FineSID combines Global–Local Quantization (GLQ) and QSCM to balance code usage and preserve semantic fidelity without artificial reassignments.The conclusion also reports consistent gains in long-tailed and large-scale scenarios.Source: Conclusions · p. 8
  • FineSID enables global gradient propagation across the codebook while preserving discrete identifiers for autoregressive generation.Source: Conclusions · p. 8

More Experiments

These experiments examine decoder-only LLM generalization, popularity bias, and SID embedding norms, alongside implementation and hyperparameter details. FineSID outperforms baselines on both Head and Tail subsets, with a larger improvement on Tail items, and enhances multilevel semantic representations.

More Experiments

  • Experiments were conducted on A100 GPUs and repeated five times with different random seeds.Source: More Experiments · p. 9

Generalization

  • The generalization study instantiates FineSID and baselines on Qwen-1.5B–Qwen-7B, with detailed Qwen-1.5B results on Toys and Beauty.The scale comparison is summarized in Table 7, and Table 6 reports Qwen-1.5B results.Source: Generalization · p. 9, Table 7 · p. 10, Table 6 · p. 10

Popularity Bias

  • FineSID outperforms all baselines on both Head and Tail subsets, with a more pronounced improvement on Tail items.Table 8 compares NDCG@5 on ML-1M and Beauty.Source: Popularity Bias · p. 9, Table 8 · p. 10

Average Norm Analysis of SID

Hyperparameters

  • The selected hyperparameters are γ = 0.6 and τ = 0.2, based on the stated candidate sets.The passage also lists candidate values for ϵ but does not state its selected value.Source: Hyperparameters · p. 10

Discussion and Analysis

The analysis presents theoretical bounds and guarantees for FineSID’s representation capacity, collapse prevention, quantization error, semantic preservation, convergence, and generalization. These results are framed around hierarchical codes, GLQ, and QSCM.

Theoretical Analysis

  • The theoretical analysis covers representation capacity, collapse prevention, and semantic preservation for GLQ and QSCM.Source: Theoretical Analysis · p. 10

Representation Capacity Analysis

  • FineSID’s hierarchical SID capacity is bounded by item entropy, code cardinality, and a Gaussian-channel term based on quantization noise.The code-cardinality bound is L log K; the analysis also identifies q as quantization error variance and H(X) as item entropy.Source: Representation Capacity Analysis · p. 11

Collapse Prevention Guarantees

  • GLQ’s collapse-prevention analysis states that soft assignment preserves nonzero gradient flow for underused SIDs and gives near-perfect utilization when τ = Θ(1/ log K).Source: Collapse Prevention Guarantees · p. 11

Quantization Error Analysis

  • FineSID’s hierarchical quantization error is bounded using per-level errors, each bounded by its codebook’s covering radius.The proof invokes residual quantization and codebook covering properties.Source: Quantization Error Analysis · p. 11

Semantic Preservation Theory

  • QSCM’s double-quantization alignment loss ensures Lipschitz continuity, and a corollary addresses preservation of pairwise collaborative similarity.The loss enforces bidirectional consistency between the quantized and original representations.Source: Semantic Preservation Theory · p. 11

Optimization Convergence Analysis

  • The unified FineSID objective is stated to converge exponentially fast, with the rate parameter depending on the loss landscape’s condition number.The bound’s constant also depends on initialization.Source: Optimization Convergence Analysis · pp. 11–12

Generalization Bounds

  • The paper states a high-probability generalization error bound and a non-trivial recall guarantee for items with frequency f ≤ 1/K.The generalization analysis uses Rademacher complexity and identifies N as the number of training sequences.Source: Generalization Bounds · p. 12

Related papers

Loading 2609.36670v2…