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

TL;DR

Rethinking Generative Recommender Tokenizer: Recsys-Native Encoding and Semantic Quantization Beyond LLMs introduces ReSID, a recommendation-native SID framework that aligns representation learning and quantization with the needs of autoregressive SID decoding. It presents Field-Aware Masked Auto-Encoding (FAMAE) for collaboration-focused item representations and Globally Aligned Orthogonal Quantization (GAOQ) for compact, prefix-robust SID sequences, all without relying on foundation models. The approach is grounded in information-theoretic principles, with task-aware embedding metrics and a three-term SID objective that balances reconstruction and prefix-conditional uncertainty. Empirical results across ten Amazon-2023 datasets show ReSID outperforming strong baselines by over 10% while achieving up to 122x tokenization cost reductions, demonstrating both effectiveness and efficiency for scalable SID-based generative recommendation.

Abstract

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.

Figures (8)

• Figure 1: Illustration of a traditional semantic-centric SID-based generative recommendation pipeline. Item representations learned from foundation models are weakly aligned with collaborative prediction, and subsequent quantization does not account for sequential predictability in SID decoding, leading to high decoding uncertainty and suboptimal recommendation performance.
• Figure 2: Overview of ReSID. FAMAE learns recommendation-sufficient field-level item representations via masked field prediction, and GAOQ discretizes them into compact, autoregressive-decoding-friendly SIDs via global alignment.
• Figure 3: Downstream R@10 at selected FAMAE training steps plotted against Metric 1 (R@10 of the target item when all fields are masked). The right y-axis shows the corresponding Metric 2 (R@10 of the target item-ID when only the item-ID field is masked). Left: Musical Instruments. Right: Baby Products.
• Figure 4: Empirical scaling behavior on Baby Products. The x-axis shows $\log_{10}(P)$, where $P$ is the number of non-embedding model parameters, and the y-axis reports NDCG@10 on the test set. For each model size, we select the checkpoint with the best validation NDCG@10 and report its corresponding test NDCG@10. We compare ReSID with TIGER and SASRec under matched backbone parameter budgets.
• Figure 5: T-SNE visualization of item embeddings learned by different methods. Left:Semantic Category Structure (Cate1)—items are colored by their category labels. Right:Behavioral Community Structure—items are colored by communities discovered via the Louvain algorithm on a weighted item--item co-occurrence graph constructed from user interaction histories.

Theorems & Definitions (2)

• Proposition 3.1: Predictive Sufficiency Proxy
• proof