Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Differentiable Semantic ID for Generative Recommendation

Junchen Fu, Xuri Ge, Alexandros Karatzoglou, Ioannis Arapakis, Suzan Verberne, Joemon M. Jose, Zhaochun Ren

TL;DR

This work tackles the objective mismatch in generative recommendation caused by freezing semantic IDs learned for content reconstruction. It introduces DIGER, which makes SIDs differentiable through Differentiable Semantic ID with Exploratory Learning (DRIL) and two uncertainty-decay strategies that gradually shift from exploration to exploitation. By enabling end-to-end gradient flow into the SID codebook, DIGER achieves consistent gains over two-stage pipelines and competitive state-of-the-art results on diverse datasets, while mitigating codebook collapse. The approach aligns the indexing objective with downstream ranking, opening a promising direction for differentiable semantic indexing in generative recommendation systems.

Abstract

Generative recommendation provides a novel paradigm in which each item is represented by a discrete semantic ID (SID) learned from rich content. Most existing methods treat SIDs as predefined and train recommenders under static indexing. In practice, SIDs are typically optimized only for content reconstruction rather than recommendation accuracy. This leads to an objective mismatch: the system optimizes an indexing loss to learn the SID and a recommendation loss for interaction prediction, but because the tokenizer is trained independently, the recommendation loss cannot update it. A natural approach is to make semantic indexing differentiable so that recommendation gradients can directly influence SID learning, but this often causes codebook collapse, where only a few codes are used. We attribute this issue to early deterministic assignments that limit codebook exploration, resulting in imbalance and unstable optimization. In this paper, we propose DIGER (Differentiable Semantic ID for Generative Recommendation), a first step toward effective differentiable semantic IDs for generative recommendation. DIGER introduces Gumbel noise to explicitly encourage early-stage exploration over codes, mitigating codebook collapse and improving code utilization. To balance exploration and convergence, we further design two uncertainty decay strategies that gradually reduce the Gumbel noise, enabling a smooth transition from early exploration to exploitation of learned SIDs. Extensive experiments on multiple public datasets demonstrate consistent improvements from differentiable semantic IDs. These results confirm the effectiveness of aligning indexing and recommendation objectives through differentiable SIDs and highlight differentiable semantic indexing as a promising research direction.

Differentiable Semantic ID for Generative Recommendation

TL;DR

This work tackles the objective mismatch in generative recommendation caused by freezing semantic IDs learned for content reconstruction. It introduces DIGER, which makes SIDs differentiable through Differentiable Semantic ID with Exploratory Learning (DRIL) and two uncertainty-decay strategies that gradually shift from exploration to exploitation. By enabling end-to-end gradient flow into the SID codebook, DIGER achieves consistent gains over two-stage pipelines and competitive state-of-the-art results on diverse datasets, while mitigating codebook collapse. The approach aligns the indexing objective with downstream ranking, opening a promising direction for differentiable semantic indexing in generative recommendation systems.

Abstract

Generative recommendation provides a novel paradigm in which each item is represented by a discrete semantic ID (SID) learned from rich content. Most existing methods treat SIDs as predefined and train recommenders under static indexing. In practice, SIDs are typically optimized only for content reconstruction rather than recommendation accuracy. This leads to an objective mismatch: the system optimizes an indexing loss to learn the SID and a recommendation loss for interaction prediction, but because the tokenizer is trained independently, the recommendation loss cannot update it. A natural approach is to make semantic indexing differentiable so that recommendation gradients can directly influence SID learning, but this often causes codebook collapse, where only a few codes are used. We attribute this issue to early deterministic assignments that limit codebook exploration, resulting in imbalance and unstable optimization. In this paper, we propose DIGER (Differentiable Semantic ID for Generative Recommendation), a first step toward effective differentiable semantic IDs for generative recommendation. DIGER introduces Gumbel noise to explicitly encourage early-stage exploration over codes, mitigating codebook collapse and improving code utilization. To balance exploration and convergence, we further design two uncertainty decay strategies that gradually reduce the Gumbel noise, enabling a smooth transition from early exploration to exploitation of learned SIDs. Extensive experiments on multiple public datasets demonstrate consistent improvements from differentiable semantic IDs. These results confirm the effectiveness of aligning indexing and recommendation objectives through differentiable SIDs and highlight differentiable semantic indexing as a promising research direction.
Paper Structure (18 sections, 3 theorems, 20 equations, 7 figures, 5 tables)

This paper contains 18 sections, 3 theorems, 20 equations, 7 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Our Method: DIGER
  5. DRIL: Differentiable Semantic ID with Exploratory Learning
  6. Two Uncertainty Decay Strategies
  7. Model Training
  8. Experimental Setup
  9. Experimental results
  10. Comparison with differentiable SID and conventional (RQ1)
  11. Comparison with SOTA baselines (RQ2)
  12. Ablation Study (RQ3)
  13. Semantic ID Dynamics and Stability (RQ4)
  14. Conclusions
  15. Theoretical Analysis
  16. ...and 3 more sections

Key Result

theorem 1

Assume the two-stage scheme first selects $\phi$ from a subset $\mathcal{A}\subseteq\Phi$ (e.g., $\mathcal{A}=\arg\min_\phi \mathcal{L}_{\mathrm{aux}}(\phi)$ or the set of outputs of a pretraining routine), and then optimizes $\theta$ with $\phi$ fixed. Define the value function $g(\phi):=\inf_{\the

Figures (7)

  • Figure 1: Conventional vs. Generative recommendation with Differentiable SID. A “brickmaker--builder” sketch is used to illustrate the tokenizer and recommender.
  • Figure 2: Comparison of DIGER and STE on Amazon Beauty Dataset. Top: validation NDCG@10 across epochs. Bottom: code balance summarized by the mean over three codebook levels' coverage, with error bars indicating the standard deviation across levels. Grey dashed vertical line marks the STE early-stop epoch.
  • Figure 3: Comparison of STE vs. DIGER. STE: Relies on deterministic hard selection (e.g., sequence 1-3-2), where gradients only backpropagate to the selected indices (Hard Update). DIGER: Introduces Gumbel noise to encourage exploratory learning. For instance, the noise may shift a deterministic assignment from code 3 to 2 (changing the path from 1-3-2 to 1-2-2), exploring alternative semantic representations. Crucially, DIGER employs Soft Update, allowing gradients to flow to all codebook weighted by their Gumbel-Softmax probabilities, while the noise level is progressively reduced via uncertainty decay strategies.
  • Figure 4: Two uncertainty decay strategies for exploration--exploitation: standard deviation--based uncertainty decay (SDUD) and frequency-based uncertainty decay (FrqUD).
  • Figure 5: Hyperparameter analysis on Beauty and Instruments. We vary the hot-threshold $r$ and $\lambda$, and report Recall@10 and NDCG@10.
  • ...and 2 more figures

Theorems & Definitions (5)

  • theorem 1: Two-stage freezing as restricted minimization over $\phi$
  • theorem 2: Arbitrary suboptimality under objective mismatch
  • Remark 1
  • theorem 3: Entropy regularization maximizes effective code usage
  • Remark 2