Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cognitive-Aligned Spatio-Temporal Large Language Models For Next Point-of-Interest Prediction

Penglong Zhai, Jie Li, Fanyi Di, Yue Liu, Yifang Yuan, Jie Huang, Peng Wu, Sicong Wang, Mingyang Yin, Tingting Hu, Yao Xu, Xin Li

TL;DR

The paper tackles next POI prediction by bridging large language models with spatio-temporal and cognitive knowledge. It introduces CoAST, a two-stage framework comprising spatio-temporal knowledge acquisition through continued pretraining on SID-tokenized POI sequences and cognitive preference alignment via SFT and Direct Preference Optimization RL. The method defines four cognitive dimensions—Temporal Consistency, Spatial Clustering, Profile Alignment, and Situational Awareness—and optimizes for cognitively plausible recommendations, achieving substantial offline gains and favorable online A/B results on AMAP. The work demonstrates practical feasibility through online deployment techniques (PD and MTP Eagle) and provides a data-driven framework for aligning LLMs with human cognition in location-based services, with significant improvements in P-CTR, U-CTR, and engagement metrics.

Abstract

The next point-of-interest (POI) recommendation task aims to predict the users' immediate next destinations based on their preferences and historical check-ins, holding significant value in location-based services. Recently, large language models (LLMs) have shown great potential in recommender systems, which treat the next POI prediction in a generative manner. However, these LLMs, pretrained primarily on vast corpora of unstructured text, lack the native understanding of structured geographical entities and sequential mobility patterns required for next POI prediction tasks. Moreover, in industrial-scale POI prediction applications, incorporating world knowledge and alignment of human cognition, such as seasons, weather conditions, holidays, and users' profiles (such as habits, occupation, and preferences), can enhance the user experience while improving recommendation performance. To address these issues, we propose CoAST (Cognitive-Aligned Spatial-Temporal LLMs), a framework employing natural language as an interface, allowing for the incorporation of world knowledge, spatio-temporal trajectory patterns, profiles, and situational information. Specifically, CoAST mainly comprises of 2 stages: (1) Recommendation Knowledge Acquisition through continued pretraining on the enriched spatial-temporal trajectory data of the desensitized users; (2) Cognitive Alignment to align cognitive judgments with human preferences using enriched training data through Supervised Fine-Tuning (SFT) and a subsequent Reinforcement Learning (RL) phase. Extensive offline experiments on various real-world datasets and online experiments deployed in "Guess Where You Go" of AMAP App homepage demonstrate the effectiveness of CoAST.

Cognitive-Aligned Spatio-Temporal Large Language Models For Next Point-of-Interest Prediction

TL;DR

The paper tackles next POI prediction by bridging large language models with spatio-temporal and cognitive knowledge. It introduces CoAST, a two-stage framework comprising spatio-temporal knowledge acquisition through continued pretraining on SID-tokenized POI sequences and cognitive preference alignment via SFT and Direct Preference Optimization RL. The method defines four cognitive dimensions—Temporal Consistency, Spatial Clustering, Profile Alignment, and Situational Awareness—and optimizes for cognitively plausible recommendations, achieving substantial offline gains and favorable online A/B results on AMAP. The work demonstrates practical feasibility through online deployment techniques (PD and MTP Eagle) and provides a data-driven framework for aligning LLMs with human cognition in location-based services, with significant improvements in P-CTR, U-CTR, and engagement metrics.

Abstract

The next point-of-interest (POI) recommendation task aims to predict the users' immediate next destinations based on their preferences and historical check-ins, holding significant value in location-based services. Recently, large language models (LLMs) have shown great potential in recommender systems, which treat the next POI prediction in a generative manner. However, these LLMs, pretrained primarily on vast corpora of unstructured text, lack the native understanding of structured geographical entities and sequential mobility patterns required for next POI prediction tasks. Moreover, in industrial-scale POI prediction applications, incorporating world knowledge and alignment of human cognition, such as seasons, weather conditions, holidays, and users' profiles (such as habits, occupation, and preferences), can enhance the user experience while improving recommendation performance. To address these issues, we propose CoAST (Cognitive-Aligned Spatial-Temporal LLMs), a framework employing natural language as an interface, allowing for the incorporation of world knowledge, spatio-temporal trajectory patterns, profiles, and situational information. Specifically, CoAST mainly comprises of 2 stages: (1) Recommendation Knowledge Acquisition through continued pretraining on the enriched spatial-temporal trajectory data of the desensitized users; (2) Cognitive Alignment to align cognitive judgments with human preferences using enriched training data through Supervised Fine-Tuning (SFT) and a subsequent Reinforcement Learning (RL) phase. Extensive offline experiments on various real-world datasets and online experiments deployed in "Guess Where You Go" of AMAP App homepage demonstrate the effectiveness of CoAST.
Paper Structure (44 sections, 7 equations, 5 figures, 6 tables)

This paper contains 44 sections, 7 equations, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Next POI Prediction
  4. LLMs for Recommendation
  5. Preliminaries
  6. Problem Formulation
  7. Cognitive-Aligned POI Prediction
  8. Temporal Consistency
  9. Spatial Clustering
  10. Profile Alignment
  11. Situational Awareness
  12. Methodologies
  13. Spatio-temporal Knowledge Acquisition
  14. Semantic ID generation
  15. Continued Pretraining on Spatio-temporal Corpus
  16. ...and 29 more sections

Figures (5)

  • Figure 1: Offline Training and Online Service Procedure of proposed method CoAST. Offline Training: (1) POIs are tokenized as Semantic IDs (SIDs); (2) a base LLM is continually pretrained on the collected corpus, yielding ST-LLM; (3) multi-level user profiles are distilled from the global tracking database with the annotations of LLM; (4) cognitive-aligned data are produced through LLM annotation and human screening; (6) ST-LLM is further fine-tuned and aligned to finish this procedure. Online Serving: CoAST is stored in a model engine and invoked through a data-service engine that converts real-time context—user profile, recent check-ins, current time, location, and weather—into an instruction prompt. The model predicts the next POI in SID form, which is mapped back to a readable location and returned to the mobile client.
  • Figure 2: Cascade ranking system of "Guess Where You Go" deployed on AMAP, which includes three stages from the left to the right: Retrieval, Ranking, and Cognitive Re-ranking.
  • Figure 3: Impact on the performance of sequence length (input tokens) and number of tokens involved in continued pretraining stage. a-CAS denotes the average Cognition Alignment score computed by averaging the four cognitive scores.
  • Figure 4: Comparison of Model Inference p99 Latency (ms) under a 50 QPS Load Test on 2$\times$ NVIDIA H20 GPUs.
  • Figure 5: Case comparison of recommendation results of CoAST and the traditional basline RSs. The top row represents the user's real-time check-ins with different action types. The middle and bottom row represent the recommendation results of traditional basline RSs and CoAST, respectively.