Elsevier

Neurocomputing

Volume 488, 1 June 2022, Pages 107-119
Neurocomputing

FG-CF: Friends-aware graph collaborative filtering for POI recommendation

https://doi.org/10.1016/j.neucom.2022.02.070Get rights and content

Abstract

Collaborative filtering approach greatly promotes the development and application of personalized recommendation. In location-based social networks (LBSNs), the sparsity of check-in data is one of the main obstacles for traditional Point-of-Interest (POI) recommendation models. Graph convolutional network (GCN) is an efficient tool to overcome this kind of problems, which enhances the representational ability of embeddings by capture high-order connectivity of users and POIs. In real applications, social tie is a crucial factor for POI recommendation that ignored in most current graph-based methods. Moreover, most message aggregation functions fail to capture contextual information. To address these problems, a novel framework named Friends-aware Graph Collaborative Filtering (FG-CF) is proposed in this paper, which incorporates social information into a user-POI graph. Firstly, a user-POI correlation matrix is estimated by check-in data and social links, and then, user embedding is updated according to the user-POI correlation matrix. Secondly, interaction messages are constructed in a novel way by integrating nodes’ ego embeddings, neighbors’ embeddings and social embeddings. Thirdly, by aggregating previous state embeddings and non-linear combination of neighbor messages with interaction messages, a new message aggregation function is present to update user and POI embeddings. Fourthly, we concatenate embeddings from each additional interaction layer to get the final embeddings, and inner product is used to compute the preference score of a user to a targeted POI. Finally, extensive experiments on two large-scale LBSN datasets demonstrate the superiority of our model over several state-of-the-art approaches.

Introduction

Currently, with the development of positioning technologies and widely used smart terminals, various applications involve social function and location information. Location-based Social Networks (LBSNs), such as Gowalla1 and Yelp2, have developed rapidly in recent years. These platforms allow users to establish cyber links with others and share experiences about geographical places they visited. These attractive places visited or checked-in by users are called Point-of-Interests (POIs), for example, shopping mall, restaurant, museum, as well as other interesting places. And recommending places to users who never visited on a LBSN platform is called POI recommendation. Fig. 1 is a graph representation of users, POIs and social links in LBSNs. As one of the hottest research topics, POI recommendation has a wide range of practical significance and application prospect.

Recent decades, many POI recommendation approaches have been proposed, in which Collaborative Filtering (CF) is one of the most important methods. CF can be categorized into memory-base CF and model-based CF. Memory-based CF infers users’ preferences to POIs by aggregating ratings from similar users or POIs. For example, user-based CF estimates users’ preferences according to users’ similarity [1], [2], [3], [4] while POI-based CF considers POIs’ similarity [1], [3], [5], [6], [7]. Memory-based CF can be easily implemented and well scaled with co-rated users or items for its low computing and memory consumption.

Despite the effectiveness of memory-based CF, it greatly suffers from data sparsity and cold-start problems. Hence, model-based CF was proposed, which combines collaborative filtering methods with advanced models, such as matrix factorization [8], [9], machine learning methods [10], [11], neural networks [12], [13], [14], [15] and so on.

Recent years, plenty of work (for example, [16], [17]) has proved that graph convolutional network (GCN) is a powerful tool to extract latent features from graph structures, which are constructed by objects. In real-world, interactions between users and items can be constructed as bipartite graphs, GCN can be used to obtain representations with structural information from complex relationships between users and items. Since GCNs can alleviate data sparsity for exploring more nodes’ information when update target node’s embedding, series of GCN-based collaborative filtering methods were proposed to improve the performance of recommender system. Most existing GCN-based CF methods [18], [19] use the bipartite graph for collaborative signals’ propagation. The core idea of GCN-based CF methods is message passing, in which process each node aggregates neighborhood embeddings to update ego embedding. By stacking multiple convolution layers, high-order connectivity representations of users and items can be extracted.

Despite effectiveness in learning representations from graph structures, most existing GCN-based CF methods still suffer from some problems: (1) Social information, which is a crucial factor in POI recommender systems, is ignored. It contributes to limited representational capability of users and POIs because friends-effect was neglected. (2) Commonly used message construction and aggregation functions in GCN-based CF fail to exploit contextual information of user and POI embeddings, so that further limit the recommendation performance.

To address aforementioned challenges, we propose a Friends-aware Graph Collaborative Filtering (FG-CF) model as demonstrated in Fig. 2. FG-CF incorporates social influence into GCN-based CF to extract not only user-POI interaction information but also social information. In FG-CF, we construct a user-POI interaction graph. Using this graph structure, we can exploit high-order similarity of users or POIs through message construction and message aggregation. To address the first problem, we update users’ embeddings by incorporating social influence. Specifically, we use the check-in matrix and the social matrix to estimate the relevance between users and POIs to get a user-POI correlation matrix. Then, we use this correlation matrix as weight matrix to update users’ embeddings by aggregating weighted neighbors’ embeddings. In message construction stage, we construct two types of messages, (a) neighbor messages and (b) interaction messages that involve neighbor embeddings and social embeddings, where social embeddings are defined in the previous step. To address the second problem, we update every node’s embedding by using a message aggregation function we design which is different from all graph-based CF methods as far as we know. FG-CF’s message aggregation function concatenates node’s previous state embedding with the next state embedding (combination of neighbor messages and interaction messages). By doing so, FG-CF can exploit more contextual information from user-POI graph. The final embeddings are obtained by concatenating embeddings from each interaction layer except the first one. After final embeddings of user and POI are obtained, we use inner product to predict users’ preference scores to POIs. In the end, we recommend POIs to a test user with top-k highest preference scores.

To make a summary, the main contributions of our work are three-folds:

  • We propose a FG-CF model, a novel graph-based POI recommendation framework which takes social influence into account. FG-CF has an outstanding performance in extracting not only interaction information, but also social information from friends.

  • We design a message construction function that incorporates social influence and a message aggregation function which jointly considers previous state embeddings and non-linear combination of neighbor messages with interaction messages.

  • We conduct extensive experiments on two public POI recommendation datasets, i.e., Gowalla and Yelp to evaluate the performance of the FG-CF model. Experimental results show that our FG-CF model outperforms several state-of-art collaborative filtering approaches to a great extent.

The remainder of our paper is organized as follows: In Section 2, we introduce some related work and discuss the relationship with our approach. In Section 3, we demonstrate the proposed approach FG-CF in detail. In Section 4, we present the results of comparative experiments, ablation experiments and parameter experiments of our method on Gowalla and Yelp datasets. In Section 5, we make a nutshell of our paper and explore some possible directions of future work.

Section snippets

Related Work

In this section, we review two kinds of side information that can be exploited for POI recommendation and summarize existing CF-based recommendation researches that can be divided into: memory-based CF, model-based CF and graph-based CF.

Problem Statement

In this section, we introduce some concepts and definitions in POI recommendation, then notations used in this paper are summarized in Table 1. Besides, we give the general paradigm of GCN-based collaborative filtering (GCN-based CF) model.

Definition 1 (POI): A POI lL is a geographical place (e.g., a restaurant, a museum or a shopping mall) that users may be interested in. A POI has two attributes, i.e., identifier and coordinates, described by a triad <POI_id,longitude,latitude>.

Definition 2

Friends-aware Graph Collaborative Filtering (FG-CF)

We now present our FG-CF model, that consisted of three components in our model: (1) a social embedding construction layer which incorporates social information into user embeddings; (2) multiple interaction layers that contain message construction, message aggregation for the purpose of embedding update; (3) a prediction layer that concatenates every additional layer’s embeddings and outputs the preference scores of users to POIs using inner product of user embeddings and POI embeddings. The

Datasets Description

We use two public POI recommendation datasets (i.e., Gowalla and Yelp) to evaluate the performance of FG-CF. There are several types of data including check-in records, social links of users and coordinates of POIs in the datasets. The detailed statistics of these two datasets are shown in Table 2.

  • (1)

    Gowalla: Gowalla dataset has 1,278,274 check-in records generated by 18,737 users and 32,510 POIs from February 2009 to October 2010.

  • (2)

    Yelp: Yelp dataset has 860,888 check-in records generated by 30887

Conclusion and Future Work

In this paper, we introduced a novel POI recommendation model, named Friends-aware Graph Collaborative Filtering (FG-CF). FG-CF incorporates social influence into user-POI bipartite graph structure to exploit not only interaction information but also social information. Meanwhile, FG-CF’s message aggregation function comprehensively considers the previous state embeddings and non-linear combination of neighbor messages and interactions messages (that can be seen as next state embeddings). This

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

This work was supported in part by the National Natural Science Foundation of China under Grants 71774159,China Postdoctoral Science Foundation under Grant 2018M642358 and 2021T140707, Fundamental Research Funds for the Central Universities of China under Grant 2015XKMS085, Sichuan Science and Technology Program under grants 2021JDJQ0021 and 22ZDYF2680, Chengdu Major Science and Technology Innovation Project under grant 2021-YF08-00156-GX, Chengdu Technology Innovation and Research and

Zhuo Cai received the bachelor’s degree in Computer Science and Technology from China University of Mining and Technology, Xuzhou, China, in 2020. He is currently pursuing the master’s degree in Computer Science and Technology from China University of Mining and Technology. His main research interests include Recommender Systems and Graph Neural Networks.

References (52)

  • J. Wasilewski et al.

    Intent-aware item-based collaborative filtering for personalised diversification

  • F. Xue et al.

    Deep item-based collaborative filtering for top-n recommendation

    ACM Transactions on Information Systems (TOIS)

    (2019)
  • Y. Koren et al.

    Matrix factorization techniques for recommender systems

    Computer

    (2009)
  • D. Lian et al.

    Geomf: joint geographical modeling and matrix factorization for point-of-interest recommendation

  • J.L. Herlocker et al.

    An algorithmic framework for performing collaborative filtering

  • G.M. Dakhel et al.

    A new collaborative filtering algorithm using k-means clustering and neighbors’ voting

  • X. He et al.

    Neural collaborative filtering

  • S. Sedhain et al.

    Autorec: Autoencoders meet collaborative filtering

  • S. Zhang et al.

    Gcn-based user representation learning for unifying robust recommendation and fraudster detection

  • X. Wang et al.

    Neural graph collaborative filtering

  • L. Chen et al.

    Revisiting graph based collaborative filtering: A linear residual graph convolutional network approach

    Proceedings of the AAAI conference on artificial intelligence

    (2020)
  • J.-D. Zhang et al.

    igslr: personalized geo-social location recommendation: a kernel density estimation approach

  • Y. Zhang et al.

    Personalized geographical influence modeling for poi recommendation

    IEEE Intelligent Systems

    (2020)
  • L. Huang et al.

    Dan-snr: A deep attentive network for social-aware next point-of-interest recommendation

    ACM Transactions on Internet Technology (TOIT)

    (2020)
  • N. Wang et al.

    A new personalized poi recommendation based on time-aware and social influence

  • B. Yang et al.

    Social collaborative filtering by trust

    IEEE transactions on pattern analysis and machine intelligence

    (2016)
  • Cited by (0)

    Zhuo Cai received the bachelor’s degree in Computer Science and Technology from China University of Mining and Technology, Xuzhou, China, in 2020. He is currently pursuing the master’s degree in Computer Science and Technology from China University of Mining and Technology. His main research interests include Recommender Systems and Graph Neural Networks.

    Guan Yuan received the B.S., M. S. and Ph.D. degrees in Computer Science and Technology from China University of Mining and Technology, Xuzhou, China, in 2004, 2009 and 2012, respectively. He is currently a Professor with the School of Computer Science and Technology, China University of Mining and Technology. He has led several research projects in the areas of Spatial-Temporal Data Mining, Recommender Systems, as well as Social Networks.

    Shaojie Qiao received the B.S. and Ph.D. degrees from Sichuan University, Chengdu, China, in 2004 and 2009, respectively. He is currently a Professor with the School of Software Engineering, Chengdu University of Information Technology, Chengdu, China. He has led several research projects in the areas of Moving Objects Databases, Trajectory Prediction, and Trajectory Data Mining.

    Song Qu received the B.S., M. S. in Computer Science and Technology from China University of Mining and Technology, Xuzhou, China, in 2004 and 2011 respectively. He is currently pursuing the Ph.D. degree in Computer Science and Technology from China University of Mining and Technology. His main research interests include Trajectory Data Mining and Moving Objects Databases.

    Yanmei Zhang received the B.S. degree in Computer Science of Computer Science from the Nanjing Tech University, in 2007 and the Ph.D. degree in Computer Science from China University of Mining and Technology, Xuzhou, China, in 2012. She is currently an Associate Professor with the School of Computer Science from China University of Mining and Technology, Xuzhou, China. Her research interests include Software Engineering and Program Analysis and Testing.

    Rui Bing received his bachelor’s degree and master’s degree at the Northwest Normal University, Lanzhou, China, in 2017 and 2020 respectively. He is currently pursuing the doctor’s degree at the School of Computer Science and Technology, China University of Mining and Technology. His main research interests include Graph Data Mining and Graph Neural Networks.

    View full text