Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Pctx: Tokenizing Personalized Context for Generative Recommendation

Qiyong Zhong, Jiajie Su, Yunshan Ma, Julian McAuley, Yupeng Hou

TL;DR

This work tackles the limitation of static, non-personalized tokenization in generative recommendation by introducing Pctx, a personalized context-aware tokenizer that tokenizes each action conditioned on a user’s history. Pctx derives rich user-context representations, condenses them into per-item centroids, fuses these with item features, and quantizes the result into multiple semantic IDs via RQ-VAE, enabling multi-facet, personalized predictions. The training procedure uses data augmentation and the inference process employs beam search to generate and aggregate multiple potential semantic IDs, improving recall and ranking metrics. Experiments on three Amazon dataset categories show up to 11.44% gains in NDCG@10 over non-personalized baselines, with ablations confirming the contribution of contextualization, clustering, redundancy merging, and multi-facet generation. Overall, Pctx advances GR by enabling user-specific interpretation of items, improving personalization and interpretability while preserving scalability.

Abstract

Generative recommendation (GR) models tokenize each action into a few discrete tokens (called semantic IDs) and autoregressively generate the next tokens as predictions, showing advantages such as memory efficiency, scalability, and the potential to unify retrieval and ranking. Despite these benefits, existing tokenization methods are static and non-personalized. They typically derive semantic IDs solely from item features, assuming a universal item similarity that overlooks user-specific perspectives. However, under the autoregressive paradigm, semantic IDs with the same prefixes always receive similar probabilities, so a single fixed mapping implicitly enforces a universal item similarity standard across all users. In practice, the same item may be interpreted differently depending on user intentions and preferences. To address this issue, we propose a personalized context-aware tokenizer that incorporates a user's historical interactions when generating semantic IDs. This design allows the same item to be tokenized into different semantic IDs under different user contexts, enabling GR models to capture multiple interpretive standards and produce more personalized predictions. Experiments on three public datasets demonstrate up to 11.44% improvement in NDCG@10 over non-personalized action tokenization baselines. Our code is available at https://github.com/YoungZ365/Pctx.

Pctx: Tokenizing Personalized Context for Generative Recommendation

TL;DR

This work tackles the limitation of static, non-personalized tokenization in generative recommendation by introducing Pctx, a personalized context-aware tokenizer that tokenizes each action conditioned on a user’s history. Pctx derives rich user-context representations, condenses them into per-item centroids, fuses these with item features, and quantizes the result into multiple semantic IDs via RQ-VAE, enabling multi-facet, personalized predictions. The training procedure uses data augmentation and the inference process employs beam search to generate and aggregate multiple potential semantic IDs, improving recall and ranking metrics. Experiments on three Amazon dataset categories show up to 11.44% gains in NDCG@10 over non-personalized baselines, with ablations confirming the contribution of contextualization, clustering, redundancy merging, and multi-facet generation. Overall, Pctx advances GR by enabling user-specific interpretation of items, improving personalization and interpretability while preserving scalability.

Abstract

Generative recommendation (GR) models tokenize each action into a few discrete tokens (called semantic IDs) and autoregressively generate the next tokens as predictions, showing advantages such as memory efficiency, scalability, and the potential to unify retrieval and ranking. Despite these benefits, existing tokenization methods are static and non-personalized. They typically derive semantic IDs solely from item features, assuming a universal item similarity that overlooks user-specific perspectives. However, under the autoregressive paradigm, semantic IDs with the same prefixes always receive similar probabilities, so a single fixed mapping implicitly enforces a universal item similarity standard across all users. In practice, the same item may be interpreted differently depending on user intentions and preferences. To address this issue, we propose a personalized context-aware tokenizer that incorporates a user's historical interactions when generating semantic IDs. This design allows the same item to be tokenized into different semantic IDs under different user contexts, enabling GR models to capture multiple interpretive standards and produce more personalized predictions. Experiments on three public datasets demonstrate up to 11.44% improvement in NDCG@10 over non-personalized action tokenization baselines. Our code is available at https://github.com/YoungZ365/Pctx.
Paper Structure (47 sections, 2 equations, 7 figures, 6 tables)

This paper contains 47 sections, 2 equations, 7 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Problem Formulation
  4. Personalized Action Tokenization
  5. Personalized Context Representation
  6. User Context Encoding.
  7. Multi-Facet Condensation of Context Representations.
  8. Personalized Semantic ID
  9. Semantic ID Construction from Context Representations.
  10. Redundant Semantic ID Merging.
  11. Generative Recommendation under Pctx
  12. Training with Data Augmentation.
  13. Multi-Facet Semantic ID Generation.
  14. Discussion
  15. Static Tokenizers.
  16. ...and 32 more sections

Figures (7)

  • Figure 1: The comparison between the current paradigm applying the static tokenizer and Pctx.
  • Figure 2: Overall framework of Pctx.
  • Figure 3: The number of personalized semantic IDs (simplified as SIDs) every item possesses.
  • Figure 4: Case Study. The upper row denotes a story-driven game player, while the lower row depicts a real-time strategy game player. The same item StarCraft II is tokenized into different semantic IDs under different user context, reflecting its multifaceted attributes. SID denotes semantic ID.
  • Figure 5: Analysis of performance (NDCG@10, $\uparrow$) w.r.t. the augmentation probability $\gamma$.
  • ...and 2 more figures