Heterogeneous graph-based joint representation learning for users and POIs in location-based social network

doi:10.1016/j.ipm.2019.102151

Information Processing & Management

Volume 57, Issue 2, March 2020, 102151

https://doi.org/10.1016/j.ipm.2019.102151 Get rights and content

Highlights

•
We propose a novel framework for representation learning in LBSN by building a heterogeneous LBSN graph. This heterogeneous graph enables a joint modeling of various contextual factors including geographical influence, social relationship and temporal information.
•
We devise a simple yet effective method to estimate three kinds of node-to-node relatedness, which jointly consider various contextual factors to better reflect the user spatial behavior and their social relationships.
•
Through extensive experiments on two publicly available datasets, we show that the proposed UP2VEC can achieve significantly improvement in POI recommendation and social link prediction.

Abstract

Learning latent representations for users and points of interests (POIs) is an important task in location-based social networks (LBSN), which could largely benefit multiple location-based services, such as POI recommendation and social link prediction. Many contextual factors, like geographical influence, user social relationship and temporal information, are available in LBSN and would be useful for this task. However, incorporating all these contextual factors for user and POI representation learning in LBSN remains challenging, due to their heterogeneous nature. Although the encouraging performance of POI recommendation and social link prediction are delivered, most of the existing representation learning methods for LBSN incorporate only one or two of these contextual factors. In this paper, we propose a novel joint representation learning framework for users and POIs in LBSN, named UP2VEC. In UP2VEC, we present a heterogeneous LBSN graph to incorporate all these aforementioned factors. Specifically, the transition probabilities between nodes inside the heterogeneous graph are derived by jointly considering these contextual factors. The latent representations of users and POIs are then learnt by matching the topological structure of the heterogeneous graph. For evaluating the effectiveness of UP2VEC, a series of experiments are conducted with two real-world datasets (Foursquare and Gowalla) in terms of POI recommendation and social link prediction. Experimental results demonstrate that the proposed UP2VEC significantly outperforms the existing state-of-the-art alternatives. Further experiment shows the superiority of UP2VEC in handling cold-start problem for POI recommendation.

Introduction

With the prevalence of mobile devices and the fast development of wireless networks, location-based service such as Foursquare¹ and Gowalla² are widely used worldwide. Unlike general social networks, the users in these location-based services can share the focused geographic entities (i.e., point of interest) they just visit and the related activities to their friends in terms of check-in. Such social network that combines social network and check-in service is called location-based social network (LBSN) (Zheng, 2011). As of 2017, Foursquare had accumulated over 12 billion Foursquare check-ins worldwide.³ That is, these LBSN services play an increasingly important role in people's daily life. Nowadays, many efforts are dedicated to exploit the huge number of check-in records to improve LBNS services. One of the findings is that the learning latent representations of user and POI can be used to enhance POI recommendation (Wang, Tan, Zhang & You, 2016; Pálovics et al., 2017) and social link prediction (Martínez, Berzal & Talavera, 2017; Wu, Zhang & Ren, 2017).

Besides check-in records made by the massive users (Zhao, Cong, Yuan & Zhu, 2015; Zhang & Chow, 2015; Bayrak & Polat, 2014), there are various contextual factors available in LBSN (Valverde-Rebaza, Roche, Poncelet & de Andrade Lopes, 2018), such as geographical influence (Li, Cong, Li, Pham & Krishnaswamy, 2015; Liu, Wei, Sun & Miao, 2014; Cranshaw, Toch, Hong, Kittur & Sadeh, 2010; Scellato, Noulas & Mascolo, 2011), social connections (Bayrak & Polat, 2014; Li, Ge, Hong & Zhu, 2016; Tang, Hu, Gao & Liu, 2013; Valverde-Rebaza, Roche, Poncelet & de Andrade Lopes, 2016), temporal information (Yuan, Cong, Ma, Sun & Thalmann, 2013; Zhao, Zhao, Yang, Lyu & King, 2016). Most of the existing user and POI representation learning methods (Li et al., 2015; Zhao, Fan, Wen & Chang, 2018; Feng et al., 2015; Xie et al., 2016) for LBSN only utilize some of these factors, resulting a partial encoding for the users and POIs towards the understanding of user spatial behavior and their social relationship. For instance, (Zhao, Fan et al., 2018) models chronological check-in sequence and social information without geographical information. (Li et al., 2015; Feng et al., 2015; Xie et al., 2016) exclude social connections for representation learning. And the techniques presented in Li et al. (2015), Zhao, Fan et al. (2018) and Feng et al. (2015) only exploit check-in information, which can not handle the cold-start problem. Although these methods have delivered effective representation learning for the users and POIs in LBSN in terms of POI recommendation and social link prediction, an important problem is not well considered, i.e., how to jointly model all these contextual factors to derive the user and POI representations. In other words, encoding users and POIs in LBSN by jointly considering these contextual factors would result in a more comprehensive understanding of user spatial behavior and their social relationship, leading to better LBSN services, such as POI recommendation and social link prediction.

To this end, in this paper, we propose a heterogeneous LBSN graph-based joint representation learning framework, named as UP2VEC (namely User and POI toVector). UP2VEC takes the users and POIs, check-in records and the auxiliary social network of a LBSN as input, and generates latent representations for each user and POI as output. In UP2VEC, we first build a heterogeneous graph to include the users and POIs, and the interactions between them. This heterogeneous graph could easily facilitate the incorporation of various contextual factors into the node-to-node relatedness measure. Specifically, there are three kinds of node-to-node relatedness, namely user-POI, POI-POI and user-user. We derive the user-POI relatedness by considering both the check-in records of this user-POI pair and the geographical distance between them. The similar procedure is also applied for POI-POI relatedness measure by considering the POI to POI sequential patterns and their geographical distance. As to user-user relations, we calculate their relatedness in terms of both common check-in behaviors and the geographical distance. To model the temporal information, we further consider the time point of the POI visit when building the heterogeneous graph. Inspired by recent progress in network embedding (Zhao, Zhao et al., 2016; Zhao, Fan et al., 2018), we then adopt the effective network embedding technique (i.e., Node2Vec) to learn the latent representations for each user and POI based on the heterogeneous graph. Note that the geographical influence is modeled in all three node-to-node relations. This offers us a straightforward strategy to accommodate the cold-start users and POIs for representation learning.

We evaluate the proposed UP2VEC over two real-world datasets. The experimental results demonstrate that UP2VEC substantially outperforms the existing state-of-the-art techniques in terms of POI recommendation and social link prediction. We further show that UP2VEC provides superior ability in handling cold-start users and POIs. To sum up, the main contributions of this paper are as follows:

•
We propose a novel framework for representation learning in LBSN By building a heterogeneous LBSN graph. This heterogeneous graph enables a joint modeling of various contextual factors including geographical influence, social relationship and temporal information.
•
We devise a simple yet effective method to estimate three kinds of node-to-node relatedness, which jointly consider various contextual factors to better reflect the user spatial behavior and their social relationships.
•
Through extensive experiments on two publicly available datasets, we show that the proposed UP2VEC can achieve significantly improvement in POI recommendation and social link prediction.

The rest of this paper is organized as follows. Section 2 presents literature review for related works in POI recommendation, link prediction and network embedding. Section 3 presents key definitions and problem formulation of node representation learning, POI recommendation and link prediction. Section 4 details the proposed framework and key modules of the framework. Section 5 presents experimental evaluation in terms of POI recommendation, link prediction and parameter sensitivity analysis. Section 6 presents conclusion and future research directions.

Section snippets

Related work

Our work is mainly related to the research studies of POI recommendation, link prediction and network embedding. Therefore, we review the related literature in this section.

Problem formulation

To ease the understanding of the proposed UP2VEC, in this section, we first define the key concepts used in this paper. The problem formulation for POI recommendation and social link prediction are also presented here.

Definition 1

(Point of Interest (POI)) A POI l is defined as a unique geographical entity (e.g., a restaurant or a cinema) whose geographical coordinate is available. The collection of POIs available in a LBSN is denoted as L.

Definition 2

(Check-in) A check-in is a triplet $c = (l, u, t)$ which indicates that

Proposed framework

We need to learn the representations for LBSN users and POIs (including cold-start users and POIs) from the check-ins, social connections, and auxiliary geographical and temporal information. Therefore, we need to figure out a proper way to integrate them into a unified learning model. In LBSN, there are two main entities available, namely users and POIs. The users are connected by social connections, while POIs are connected in the check-in sequences. Users and POIs are also connected via the

Performance evaluation

In this section, we conduct extensive experiments over two real-world datasets to evaluate the efficacy of the proposed UP2VEC framework in terms of three relevant tasks: POI recommendation, cold-start POI recommendation and link prediction. Then, we also present the study of parameter sensitivity for UP2VEC.

Conclusion

In this paper, we have proposed a heterogeneous graph-based joint representation learning framework (UP2VEC) for users and POIs in LBSN. The learned representations of users and POIs are task independent. Therefore, they can be directly used as input for various LBSN tasks, such as POI recommendation and social link prediction. UP2VEC jointly models social relationships, geographical influence, and temporal information through constructing a user-POI heterogeneous graph. Besides that, it

Acknowledgments

This work was supported by the National Key R&D Program of China (No.2016YFB0801303, 2016QY01W0105), the National Natural Science Foundation of China (No.U1636219, U1804263, 61872278, U1736214), and the Plan for Scientific Innovation Talent of Henan Province (No. 184200510018).

References (52)

Q. Li et al.
Mining user similarity based on location history
R. Pálovics et al.
Location-aware online learning for top-k recommendation
Pervasive and Mobile Computing
(2017)
J.C. Valverde-Rebaza et al.
The role of location and social strength for friendship prediction in location-based social networks
Information Processing and Management
(2018)
J.-H. Wu et al.
A balanced modularity maximization link prediction model in social networks
Information Processing and Management
(2017)
L. Backstrom et al.
Find me if you can: Improving geographical prediction with social and spatial proximity
A.E. Bayrak et al.
Contextual feature analysis to improve link prediction for location based social networks
H. Chen et al.
A tutorial on network embeddings
C. Cheng et al.
Fused matrix factorization with geographical and social influence in location-based social networks
E. Cho et al.
Friendship and mobility: User movement in location-based social networks
J. Cranshaw et al.
Bridging the gap between physical location and online social networks

S. Feng et al.

Personalized ranking metric embedding for next new poi recommendation

H. Gao et al.

gSCorr: Modeling geo-social correlations for new check-ins on location-based social networks

A. Grover et al.

Node2Vec: Scalable feature learning for networks

W. Hamilton et al.

Inductive representation learning on large graphs

Advances in Neural Information Processing Systems

(2017)

X. He et al.

Neural collaborative filtering

H. Li et al.

Point-of-Interest recommendations: Learning potential check-ins from friends

R. Li et al.

Towards social user profiling: Unified and discriminative influence model for inferring home locations

X. Li et al.

Rank-GeoFM: A ranking based geographical factorization method for point of interest recommendation

D. Lian et al.

GeoMF: Joint geographical modeling and matrix factorization for point-of-interest recommendation

Y. Liu et al.

Exploiting geographical neighborhood characteristics for location recommendation

H. Ma et al.

Recommender systems with social regularization

V. Martínez et al.

A survey of link prediction in complex networks

ACM Computing Surveys

(2017)

T. Mikolov et al.

Distributed representations of words and phrases and their compositionality

Advances in Neural Information Processing Systems

(2013)

A. Noulas et al.

A random walk around the city: New venue recommendation in location-based social networks

B. Perozzi et al.

DeepWalk: Online learning of social representations

H. Pham et al.

Towards integrating real-world spatiotemporal data with social networks

Cited by (59)

Towards human-like perception: Learning structural causal model in heterogeneous graph
2024, Information Processing and Management
Heterogeneous graph neural networks have become popular in various domains. However, their generalizability and interpretability are limited due to the discrepancy between their inherent inference flows and human reasoning logic or underlying causal relationships for the learning problem. This study introduces a novel solution, HG-SCM (Heterogeneous Graph as Structural Causal Model). It can mimic the human perception and decision process through two key steps: constructing intelligible variables based on semantics derived from the graph schema and automatically learning task-level causal relationships among these variables by incorporating advanced causal discovery techniques. We compared HG-SCM to seven state-of-the-art baseline models on three real-world datasets, under three distinct and ubiquitous out-of-distribution settings. HG-SCM achieved the highest average performance rank with minimal standard deviation, substantiating its effectiveness and superiority in terms of both predictive power and generalizability. Additionally, the visualization and analysis of the auto-learned causal diagrams for the three tasks aligned well with domain knowledge and human cognition, demonstrating prominent interpretability. HG-SCM’s human-like nature and its enhanced generalizability and interpretability make it a promising solution for special scenarios where transparency and trustworthiness are paramount.
Meta-learning adaptation network for few-shot link prediction in heterogeneous social networks
2023, Information Processing and Management
Link prediction, which aims to predict future or missing links among nodes, is a crucial research problem in social network analysis. A unique few-shot challenge is link prediction on newly emerged link types without sufficient verification information in heterogeneous social networks, such as commodity recommendation on new categories. Most of current approaches for link prediction rely heavily on sufficient verified link samples, and almost ignore the shared knowledge between different link types. Hence, they tend to suffer from data scarcity in heterogeneous social networks and fail to handle newly emerged link types where has no sufficient verified link samples. To overcome this challenge, we propose a model based on meta-learning, called the meta-learning adaptation network (MLAN), which acquires transferable knowledge from historical link types to improve the prediction performance on newly emerged link types. MLAN consists of three main components: a subtask slicer, a meta migrator, and an adaptive predictor. The subtask slicer is responsible for generating community subtasks for the link prediction on historical link types. Subsequently, the meta migrator simultaneously completes multiple community subtasks from different link types to acquire transferable subtask-shared knowledge. Finally, the adaptive predictor employs the parameters of the meta migrator to fuse the subtask-shared knowledge from different community subtasks and learn the task-specific knowledge of newly emerged link types. Experimental results conducted on real-world social media datasets prove that our proposed MLAN outperforms state-of-the-art models in few-shot link prediction in heterogeneous social networks.
Improving the spatial–temporal aware attention network with dynamic trajectory graph learning for next Point-Of-Interest recommendation
2023, Information Processing and Management
Next Point-Of-Interest (POI) recommendation aim to predict users’ next visits by mining their movement patterns. Existing works attempt to extract spatial–temporal relationships from historical check-ins; however, the following critical factors have not been adequately considered: (1) structured features implied in trajectory that reflect individual visit tendency; (2) collaborative signals from other users and (3) dynamic user preference. To this end, we jointly take into full consideration the graph-structured information as well as sequential effects of user trajectory sequences and propose the Trajectory Graph enhanced Spatial–Temporal aware Attention Network (TGSTAN). Given the general preference among users and the shifts of individual interests over time, we present a novel trajectory-aware dynamic graph convolution network module (TDGCN) to facilitate the capturing of local spatial correlations. Specifically, TDGCN dynamically adjusts the normalized adjacency matrix of the trajectory graph by element-wise multiplication with self-attentive POI representations. The local trajectory graph is generated from the same training batch to reflect real-time and collaborative signals, while also following causality. Moreover, we explicitly integrate spatial–temporal interval information with bilinear interpolation to comprehensively attach relative proximity to attention mechanism when capturing long-term dependence. Extensive experiments on three real-world Location-Based Social Networks datasets (Foursquare_TKY, Weeplaces and Gowalla_CA) demonstrate that the proposed TGSTAN consistently outperforms the existing state-of-the-art baselines with an average of 8.18%, 6.59%, and 9.60% improvement on the three datasets, respectively.
Heterogeneous question answering community detection based on graph neural network
2023, Information Sciences
Topic-based communities have gradually become a considerable medium for netizens to disseminate and acquire knowledge. These communities consist of entities (actual objects, e.g., a real answer or an actual question) with different types (users, questions and answers) and are usually hidden and overlapping. Nowadays, prevalent community question answering (CQA) platforms have formed mature communities by manually marked topics and extensive accumulated user behavior. However, the ever-growing various entities and complex overlapping topic communities make it inefficient to manually label entity tags (e.g., Question labels supplement domain features; Potential user tags indicate the user's specialty.). Therefore, there is an urgent need for a mechanism that automatically finds hidden semantic communities from user social behavior and lays a foundation for community construction and intelligent recommendation of QA platforms. In this paper, we propose a Heterogeneous Community Detection Approach Based on Graph Neural Network, called HCDBG, to detect heterogeneous communities in CQA. Firstly, we define entity relationships based on user interaction behavior and employ a heterogeneous information network to uniformly represent all connections. Afterward, we exploit the heterogeneous graph neural network to fuse content and topological features of nodes for graph embedding. Finally, we convert the community detection issue in CQA into an entity clustering task in the heterogeneous information network and improve the k-means method to achieve heterogeneous community detection. Based on our knowledge of the existing literature, it is an innovative research direction that utilizes the heterogeneous graph neural network to facilitate QA community detection. Extensive experiments on authentic question-answering datasets illustrate that HCDBG outperforms baseline methods in heterogeneous community detection.
GLAE: A graph-learnable auto-encoder for single-cell RNA-seq analysis
2023, Information Sciences
Citation Excerpt :
Learning and exploiting unknown relations has become an essential issue for scRNA-seq-based cell clustering. Graph neural networks (GNNs) can make good use of the sample relation graph by passing messages over it to aggregate information and have achieved good results in many fields, such as recommendation systems [18–21], social networks [22,23], fake news detection [24], disease-gene association identification [25] and biomedical [26]. Thus, some methods such as scGNN [27] introduce GNN into the field of scRNA-seq analysis.
Single-cell RNA sequencing (scRNA-seq) methods based on graph neural networks (GNNs) can make good use of cell relation graphs. Considering the cell relation graph is unknown in most situations, some GNN-based methods generate a pre-fixed cell relation graph using all the features (i.e., genes) from a single perspective and input it into GNN models. However, these GNN-based models can be severely hurt by the pre-fixed relation graph especially when it is not well pre-obtained due to the scRNA-seq errors. In addition, such methods learn the cell relation graph from a single perspective using all the features, which ignores the different influences of different gene subsets on cell relations. In this paper, we propose a novel end-to-end GNN-based scRNA-seq method called GLAE to address the above shortcomings, which is capable of learning cell relation graphs from different perspectives adaptively during the training process. We compare GLAE with several recently proposed methods and the results on six scRNA-seq datasets show that GLAE outperforms most of the methods on clustering tasks and is able to learn a meaningful cell relation graph for downstream tasks.
A multiview graph collaborative filtering by incorporating homogeneous and heterogeneous signals
2022, Information Processing and Management
Citation Excerpt :
However, associations among heterogeneous attributes can also distinguish different interests of users, which help extend the semantic representation of users and items. By considering geographical influence, social relationship and temporal information in location-based social networks, Qiao et al. (2020) designed a heterogeneous graph to learn the representations of users and points of interests for making social link prediction. An attribute graph can be used to model the association of different attributes and help learn a rich attribute embedding of nodes.
In the industrial e-commerce recommender systems, the sparsity of user–item interaction limits the improvement of the performance of collaborative filtering recommendation. Some studies have leveraged attribute co-occurrence or similar neighbors to enhance the semantic representation quality of users and items. Previous methods consider collaborative signals of homogeneous type nodes, such as $< u s e r, u s e r > \to u s e r$ and $< i t e m, i t e m > \to i t e m$ . By exploiting homogeneous and heterogeneous signals of attribute and neighbor views, we design a multiview graph collaborative filtering (MVGCF) network for recommendation. The MVGCF model utilizes both co-occurrence features of various attribute values and collaborative preference of various neighbors to learn the embedding representation of nodes. Experimental results show that the MVGCF is superior to the state-of-the-art models in AUC and logloss metrics by 1.41% and 3.12% for MovieLens 1M dataset, and by 2.35% and 2.31% for BookCrossing dataset. Aiming at the sparse problem with a small amount of interaction records, our findings is that attribute co-occurrence and neighbor collaboration can improve the accuracy and provide a good explanation for e-commerce recommender systems.

View all citing articles on Scopus

View full text

Heterogeneous graph-based joint representation learning for users and POIs in location-based social network

Highlights

Abstract

Introduction

Section snippets

Related work

Problem formulation

Proposed framework

Performance evaluation

Conclusion

Acknowledgments

Pervasive and Mobile Computing

Information Processing and Management

Information Processing and Management

Find me if you can: Improving geographical prediction with social and spatial proximity

Contextual feature analysis to improve link prediction for location based social networks

A tutorial on network embeddings

Fused matrix factorization with geographical and social influence in location-based social networks

Friendship and mobility: User movement in location-based social networks

Bridging the gap between physical location and online social networks

Personalized ranking metric embedding for next new poi recommendation

gSCorr: Modeling geo-social correlations for new check-ins on location-based social networks

Node2Vec: Scalable feature learning for networks

Inductive representation learning on large graphs

Advances in Neural Information Processing Systems

Neural collaborative filtering

Point-of-Interest recommendations: Learning potential check-ins from friends

Towards social user profiling: Unified and discriminative influence model for inferring home locations

Rank-GeoFM: A ranking based geographical factorization method for point of interest recommendation

GeoMF: Joint geographical modeling and matrix factorization for point-of-interest recommendation

Exploiting geographical neighborhood characteristics for location recommendation

Recommender systems with social regularization

A survey of link prediction in complex networks

ACM Computing Surveys

Distributed representations of words and phrases and their compositionality

Advances in Neural Information Processing Systems

A random walk around the city: New venue recommendation in location-based social networks

DeepWalk: Online learning of social representations

Towards integrating real-world spatiotemporal data with social networks