A hierarchical point-of-interest category representation learning method based on space-time cues

By constructing a structured spatiotemporal cueing and multi-head attention mechanism, combined with a large language model and hierarchical structure, the problem of insufficient spatiotemporal guidance and hierarchical modeling in the learning of interest point category representation is solved, and more accurate category transfer prediction and more expressive representation are achieved.

CN122196538APending Publication Date: 2026-06-12CHINA UNIV OF MINING & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
CHINA UNIV OF MINING & TECH
Filing Date
2026-03-11
Publication Date
2026-06-12

AI Technical Summary

Technical Problem

Existing technologies lack a unified representation learning framework that can effectively integrate spatiotemporal statistical priors, general knowledge of large language models, and hierarchical constraints, resulting in a lack of spatiotemporal guidance in interest point category representation learning and a lack of semantic propagation flexibility in hierarchical modeling.

Method used

By acquiring user interest point category check-in sequences and hierarchical structure data, spatiotemporal distribution features are extracted and structured spatiotemporal prompts are constructed. Initial representations are generated using a large language model, and deep semantic fusion is performed by combining a multi-head attention mechanism. Finally, the interest point category representations are generated through joint optimization by mask category prediction and next category prediction tasks.

Benefits of technology

It realizes spatiotemporal semantic initialization and hierarchical semantic propagation of interest point category representation, improves the accuracy of category transition probability prediction and the generalization ability of representation, solves the long-tail category and cold start problems, and enhances the robustness and expressiveness of representation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122196538A_ABST
    Figure CN122196538A_ABST
Patent Text Reader

Abstract

The present application relates to the field of urban computing and data mining, and particularly relates to a hierarchical point-of-interest category representation learning method based on spatiotemporal cues, which comprises the following steps: obtaining a user check-in sequence and a category hierarchy, dividing a category sub-sequence and constructing a category transition pair; extracting a peak period and a spatial hotspot of each category, encoding into a spatiotemporal cue input large language model, and obtaining an initialized representation after dimension alignment; constructing an ancestor feature matrix according to the hierarchy, fusing the initialized representation and the ancestor matrix by using multi-head attention, and obtaining a hierarchical enhanced representation; jointly optimizing the hierarchical enhanced category representation through two pre-training tasks of mask category prediction and next category prediction, and obtaining a final point-of-interest category representation; and finally, predicting a next check-in category based on the learned representation. The present application combines spatiotemporal prior and hierarchical semantics, solves the long-tail category and cold start problems, improves the generalization ability of the representation, and can be used for point-of-interest recommendation, travel prediction and other tasks.
Need to check novelty before this filing date? Find Prior Art