Point-of-interest recommendation based on the spatial-temporal graph

• Date: 14 November 2024
• Pages: 585-602
• DOI: https://doi.org/10.1108/IJWIS-01-2024-0016
• Published date: 14 November 2024
• Subject Matter: Information & knowledge management,Information & communications technology,Information systems,Library & information science,Information behaviour & retrieval,Metadata,Internet
• Author: Mengyue Li,Fei Li,Zhanquan Wang

Point-of-interest recommendation

based on the spatial-temporal graph

Mengyue Li,Fei Li and Zhanquan Wang

East China University of Science and Technology, Shanghai, China

Abstract

Purpose –Point-of-interest (POI) recommendation techniques play a crucial role in mitigating information

overload and delivering tailored services. To address limitations in conventional POI recommendation

systems, constrainedby sparse user-POI interactions and incomplete considerationof temporal dynamics, POI

recommendationbased on the spatial-temporal graph (STG-POI)is proposed.

Design/methodology/approach –Spatial-temporal sequencegraphs from geographical locations and user

interaction historydata are constructed, which are used to mine spatial-temporal sequence information.Using

the data f‌iltered by the band-pass f‌ilter, graph neural networks with distance-awareness and sequence-

awareness are appliedto capture high-order spatial-temporal connectionswithin diverse graph topologies. The

model leverages contrastive learningfor self-supervised disentanglement of graph representations,providing

self-supervised signalsfor sequential and geographical intent perception, thereby achievingmore precise POI

personalization.

Findings –Compared to the baseline model GSTN, experimentson the Foursquare and Gowalla data sets

reveal that STG-POIimproves testing AUC by 2.0%, 2.1%, 2.0% and decreaseslogloss by 1.9%, 3.3%, 0.3%,

respectively. These results indicate the model’s effectiveness in capturing spatial-temporal information,

surpassingmainstream POI recommendation baseline models.

Originality/value –This approach constructsa dual graph from user interactiondata, harnessing sequential

and geographical information as self-supervised signals. It yields decoupled representations of these

inf‌luences, offering a comprehensiveinsight into user behaviors and preferences within location-based social

networks, thus enhancing recommendation accuracy and interpretability. This approach addresses the

challenge in graph convolutional network where only rough and smooth features are conducive to

recommendation by using band-pass f‌ilters to signif‌icantly reduce computational complexity, thereby

enhancing recommendation speed by f‌iltering out noise data that does not contribute to recommendation

performance. Experimental results indicate that this model surpasses currentmainstream approaches in POI

recommendationtasks, effectively integrating bothgeographical and temporal features.

Keywords Graph neural networks, POI recommendation, Spatial-temporal information

Paper type Research paper

1. Introduction

Advancements in internet and mobile positioning technologies have propelled location-

based social networks (LBSNs) to the forefront of computational geography, particularly

enhancing point-of-interest (POI) recommendation systems. Central to location-based

services (LBSs), POI recommendation facilitates a range of activities, from geo-targeted

advertising to dining bookings, underpinning platforms such as Foursquare and Google

Maps. POI recommendation encompasses a diverse array of geographical entities, such as

restaurants, hotelsand tourist sites, representing a critical dimension in spatial informatics.

The collaborative f‌iltering (CF) algorithm (Ye et al., 2010) represents a foundational

approach in recommendation systems. It segments user groups, leveraging historical

interaction data to infer user check-in preferences,thereby recommending POIs. In LBSNs,

CF presupposes similarities among users or between POIs, using these correlations to

International

Journal of Web

Information

Systems

585

Received18 January 2024

Revised19 March2024

5A pril2024

Accepted 8 April 2024

InternationalJournal of Web

InformationSystems

Vol.20 No. 6, 2024

pp. 585-602

© Emerald Publishing Limited

1744-0084

DOI 10.1108/IJWIS-01-2024-0016

The current issue and full text archive of this journal is available on Emerald Insight at:

https://www.emerald.com/insight/1744-0084.htm

facilitate location recommendations.This method capitalizes on collective behavior patterns

to enhance the relevance of POI suggestions. Contemporary research on POI

recommendation systems transcends traditional methods by fusing geographical data with

user check-in chronologies, aiming to ref‌ine system eff‌icacy. Models such as LSTPM (Sun

et al.,2020) integrate users’enduring interests with their immediate geographical

predilections, using geographic aff‌inity to model sequential behavior. Despite the

multifaceted advantagesof these POI recommendation approaches, few collaborative signals

have been identif‌ied. In addition, the user–POI interaction matrix, unlike the binary user–

item matrix in standard recommendations, exhibits high sparsity in POI systems. This

sparsity arises as users interact with only a subset of POIs within constrained periods,

amplifying data scarcityand exploration costs in the POI landscape (Fang et al.,2023).

Given the complex graph structures inherent in user–POI interactions, research

increasingly focuses on exploiting these rich structures to unearth higher-order connections

among POIs. Graph neuralnetworks (GNNs) have been extensively used in recommendation

systems for more effectively capturing relationships between higher-order neighbors. Xie

et al. (2016a) applied graph embedding techniques to POI recommendation tasks and

introduced a graph-based metric embedding model. This approach embeds POIs into a low-

dimensional space, capturing their representations eff‌iciently and dynamically tracking user

preferences. It also harmonizes user behavioral preferences with sequential inf‌luences

through POI embeddings. Xie et al. (2016b) developed a universal graph-based embedding

model, GE, centered on a POI–POI graph. In this graph, the correlations between two

consecutive check-in POIs (i.e. edge weights) are quantif‌ied using temporal intervals,

providing a nuanced understanding of POI dynamics. However, despite providing valuable

insights into organizing interaction data into graph structures for POI recommendations,

these models often underemphasize geographical and sequential relationships, which is a

crucial aspect of POI recommendations.

Therefore, this study introduces a novelapproach for POI recommendations, denoted by

the spatial-temporal graph-based POI recommendation (STG-POI) methodology. This

comprehensive framework is structured around three pivotal components. STG-POI

involves the construction of two distinct yet interconnected POI graphs: one spatial graph,

which is formulated based on geographicalproximities, and a sequential graph, crafted from

the historical interactions of users. This bifurcation allows for the nuanced exploration of

both spatial and temporal dimensions of user behavior. The methodology uses advanced

GNNs that are sensitive to both distance and sequence. These networks adeptly capture the

complex, high-order relationships inherent within the diverse topologies of the constructed

graphs. To enhance the eff‌icacy of this process, a band-pass f‌ilter is incorporated to

meticulously sieve throughthe data, effectively eliminating extraneous noise while retaining

both the subtle and pronounced features that are pivotal for accurate recommendation

systems. Finally, the STG-POI approach uses contrastivelearning techniques to achieve self-

supervised disentanglement of graph representations. This innovative strategy enables the

model to disentangle and recombine the sequential and geographical inf‌luences on user

preferences, thereby facilitating the generation of personalized recommendations. Through

this tripartite framework,the STG-POI method promises to offer a sophisticatedand nuanced

tool for POI recommendations, capturing the intricate interplay of spatial and temporal

factors that inf‌luenceuser choices.

This approach aims to strike a better balance between geographical and sequential

information, therebyenhancing the accuracy and interpretability of POI recommendations.

In summary,this study makes the following primary contributions:

IJWIS

20,6

586