LLM-Aligned Geographic Item Tokenization for Local-Life Recommendation

TL;DR

This paper tackles the geographic misalignment in text-based item tokenization for local-life recommendations by introducing LGSID, a two-stage framework that first aligns LLMs to geographic knowledge via a geography-aware reward model and a novel G-DPO algorithm, then compresses semantic representations through Hierarchical Geographic Item Tokenization. The RL-based alignment injects real-world spatial proximity into LLM embeddings, while the HGIT stage produces geography-aware discrete tokens with reconstruction-driven and entropy-regularized training. Empirical results on a large Kuaishou local-life dataset show consistent improvements in both discriminative (AUC gains up to +0.0417 on several backbones) and generative (NDCG/Hit gains up to ~+0.13) settings, supported by ablations, visualizations, and case studies that illustrate enhanced geographic awareness without sacrificing semantic fidelity. The method offers scalable, geography-conscious tokenization suitable for industrial deployment, improving delivery-distance relevance and exposure fairness in spatially constrained recommendations, with a tunable balance between semantic and geographic signals via the similarity weight in G-DPO.

Abstract

Recent advances in Large Language Models (LLMs) have enhanced text-based recommendation by enriching traditional ID-based methods with semantic generalization capabilities. Text-based methods typically encode item textual information via prompt design and generate discrete semantic IDs through item tokenization. However, in domain-specific tasks such as local-life services, simply injecting location information into prompts fails to capture fine-grained spatial characteristics and real-world distance awareness among items. To address this, we propose LGSID, an LLM-Aligned Geographic Item Tokenization Framework for Local-life Recommendation. This framework consists of two key components: (1) RL-based Geographic LLM Alignment, and (2) Hierarchical Geographic Item Tokenization. In the RL-based alignment module, we initially train a list-wise reward model to capture real-world spatial relationships among items. We then introduce a novel G-DPO algorithm that uses pre-trained reward model to inject generalized spatial knowledge and collaborative signals into LLMs while preserving their semantic understanding. Furthermore, we propose a hierarchical geographic item tokenization strategy, where primary tokens are derived from discrete spatial and content attributes, and residual tokens are refined using the aligned LLM's geographic representation vectors. Extensive experiments on real-world Kuaishou industry datasets show that LGSID consistently outperforms state-of-the-art discriminative and generative recommendation models. Ablation studies, visualizations, and case studies further validate its effectiveness.

Figures (10)

• Figure 1: Illustration of challenges of text-based methods in local-life recommendation. Without geographic awareness, the system may recommend semantically relevant items that are inaccessible to users due to long distance.
• Figure 2: Model structure of LGSID. The upper part illustrates the pipeline of RL-based Geographic LLM Alignment, while the lower part depicts the pipeline of Hierarchical Geographic Item Tokenization.
• Figure 3: T-SNE visualization of items around cluster centroids across tokenization methods.
• Figure 4: Token quantile percentiles across hierarchical levels for local-life items.
• Figure 5: Hierarchical category frequency distribution of LGSID for different SID prefixes (Aligned vs Unaligned).