Location Based Recommendation for Mobile Users Using Language Model and Skyline Query

Qiang Pu, Ahmed Lbath, Daqing HeInternational Journal of Information Technology and Computer Science 2012 · DOI

Abstract

Location based personalized recommendation has been introduced for the purpose of providing a mobile user with interesting information by distinguishing his preference and location. In most cases, mobile user usually does not provide all attributes of his preference or query. In extreme case, especially when mobile user is moving, he even does not provide any preference or query.

Introduction

Mobile users are often "on the go," making it cumbersome to type out detailed preferences or search queries. Traditional recommendation systems face two major hurdles: the Attribute Gap, where relational databases struggle with "soft" semantic attributes, and the Curse of Dimensionality, where multi-attribute skyline queries become computationally prohibitive as more criteria are added.

This paper proposes the LRS-SQIR system, which aims to bridge these gaps by understanding a user’s implicit preferences based on their history and providing fast, relevant suggestions even when the user provides zero explicit input.

Methodology: The Two-Circle Concept

The core of the LRS-SQIR system is the Two-Circle Concept, which balances physical constraints with semantic interests:

  1. Physical Position Circle (PPC): This is formed using a skyline query on "hard" attributes such as distance and price. It identifies locations that are objectively optimal choices based on proximity and cost.
  2. Virtual Preference Circle (VRC): This is formed using Information Retrieval (IR) techniques. The system estimates an "implicit query" from the user’s visiting history and ranks locations based on the semantic similarity of their textual descriptions.
The Two-Circle Concept: Intersection of Physical Position and Virtual Preference

The recommendation is generated where these two circles overlap, ensuring the suggested places are both physically accessible and personally interesting.

System Architecture

The LRS-SQIR architecture is divided into online and offline components to ensure real-time performance on mobile devices:

  • Preprocessing Module (Offline): Uses semantic clustering and Independent Component Analysis (ICA) to group documents. It pre-calculates document similarity matrices within clusters to reduce the online computational load.
  • Query Processing Module (Online): Handles real-time GPS signals and user queries. It performs skyline queries for the PPC and language-model-based query expansion for the VRC.
  • Visual Module: Integrates with mapping APIs to display the final recommendations to the user visually.
LRS-SQIR System Architecture showing the interaction between modules

Evaluation and Results

The system was evaluated using the TREC Los Angeles Times (LA) Collection for semantic retrieval and a Synthetic Chengdu Dataset (1,000 to 20,000 points) for skyline scalability.

Key Findings:

  • Efficiency: The proposed Fast Semantic Relevance Model (fSRM) completed 100 queries in 11.75 seconds, significantly faster than the 611 seconds required by standard models. Each individual query takes roughly 0.1 seconds.
  • Accuracy: fSRM achieved an average precision of 27.34%, outperforming basic language models (24.58%) in the TREC evaluation.
  • Scalability: By using the skyline as a filter rather than a final answer, the system avoids the performance degradation typically seen in high-dimensional skyline queries, keeping the recommendation set small and relevant.
  • User Experience: Total system response time is approximately 2 seconds, which is well within the acceptable range for a mobile application.

Conclusion

LRS-SQIR demonstrates that combining skyline queries with language models can effectively provide personalized recommendations for mobile users without requiring explicit input. Future research will address the "cold start" problem for new users and test the system against real-world GPS trajectories in complex urban environments.

Summary updated

Abstract

Location based personalized recommendation has been introduced for the purpose of providing a mobile user with interesting information by distinguishing his preference and location. In most cases, mobile user usually does not provide all attributes of his preference or query. In extreme case, especially when mobile user is moving, he even does not provide any preference or query.
BibTeX
@article{pu2012location,
  author = {Qiang Pu and Ahmed Lbath and Daqing He},
  title = {{Location Based Recommendation for Mobile Users Using Language Model and Skyline Query}},
  journal = {International Journal of Information Technology and Computer Science},
  year = {2012},
  doi = {10.5815/ijitcs.2012.10.02},
}
Figure