A point-of-interest suggestion algorithm in Multi-source geo-social networks

Highlights

• Modeling geo-social influence from social communities.

• A latent probabilistic generative model to capture the topics in LBSNs.

• Parameters can be inferred by Gibbs sampling.

• An effective framework to fulfill top-k suggestion.

Abstract

Newly emerging location-based social network (LBSN) services provide us with new platforms to share interests and individual experience based on their activity history. The problems of data sparsity and user distrust in LBSNs create a severe challenge for traditional recommender systems. Moreover, users’ behaviors in LBSNs show an obvious spatio-temporal pattern. Valuable extra information from microblog-based social networks (MBSNs) can be utilized to improve the effectiveness of POI suggestion. In this study, we propose a latent probabilistic generative model called MTAS, which can accurately capture the underlying information in users’ words extracted from both LBSNs and MBSNs by taking into consideration the decision probability, a latent variable indicating a user’s tendency to publish a review in LBSNs or MBSNs. Then, the parameters of the MTAS model can be inferred by the Gibbs sampling method in an effective manner. Based on MTAS, we design an effective framework to fulfill the top-$k$ suggestion. Extensive experiments on two real geo-social networks show that MTAS achieves better performance than existing state-of-the-art methods.

Introduction

Recent years have witnessed a blooming of Web 2.0, positioning systems and wireless communication technologies. Location based social networks (LBSNs), such as Foursquare² and Yelp³ have become a popular application and facilitated users’ daily life. People can expand circles of friends, share their real experience and post reviews, photos and videos. As a significant tool of LBSNs, suggestion methods learn users’ preferences from their historical records and meta-data such as social relationship and review content, then suggest underlying points-of-interest (POIs) to the specific users. POI suggestion can benefit advertising agencies with an effective way of launching advertisements to the potential consumers, and improve user viscosity to LBSN service providers as well (Zhao et al., 2016a).

One of the troublesome problems in POI suggestion is data sparsity (Xie et al., 2016), especially in out-of-town areas. It stems from incompleteness of reviews posted in the LBSNs and degrades the accuracy of POI suggestion. When a user visits an area in his city or an out-of-town place which he is not familiar with, this problem becomes extremely severe and LBSNs can hardly suggest ideal results.

Meanwhile, the loose relationship of users in LBSNs make the suggestion results in out-of-town scenario trustless. Users are declined to trust friends in LBSNs since they usually have different backgrounds and interests. MBSNs (Microblog-Based Social Networks) (Xiong et al., 2018) such as Facebook and Twitter are platforms for users to communicate with close friends and share their current status. Sufficient valuable information is generated and can be used for POI suggestion. A typical application scenario is illustrated in Fig. 1. Due to the data sparsity of LBSNs, users are able to acquire extra information from Twitter, a MBSN, which can help a user make a more accurate decision. A geo-social network can be viewed as a heterogeneous information system (Bu et al., 2019, Cao et al., 2019, Li et al., 2019) composed of a LBSN and a MBSN.

Users’ preferences in LBSNs show some temporal and geographical patterns. POIs with similar attributes are usually clustered in a specific region. For instance, users prefer to visiting a street with more restaurants which can provide more choices, that is, a user’s visited places are often clustered together. A user is more likely to visit places nearby rather than the distant ones. Moreover, a user tends to visit a place at right time, e.g., he is more likely to visit a restaurant at lunch time rather than a cinema. Thus the spatio-temporal factors should be taken into account when suggesting POIs.

Inspired by these motivation, we introduce a latent probabilistic generative model called Multi-source Topic Awareness Suggestion (MTAS) to fuse the information from geo-social networks. Fig. 2 shows the basic idea of the proposed MTAS model. The geo-social network can be classified into four distinct information layers, and each layer corresponds to a part of “4W” (i.e., who, when, where and what) information structure. The content layer represents reviews from users; the social layer indicates different topologies of relationships between users in LBSNs and MBSNs, which generates different kinds of communities; the geographical layer and temporal layer show the spatio-temporal information of a posting event. The behavior of posting reviews about POIs bridges the gap between online social networks and offline physical world. Given a review, a layer link connects these four layers. An anchor link connects two different accounts of one user in a LBSN and a MBSN, respectively.

In order to study the applicability of MTAS, we explore its working mechanism and performance in two scenarios: (1) home-town suggestion is supposed that a specific user is located at his familiar regions such as his living or working areas. He appeals to suggestion services of LBSNs for new suggestion based on his activity records. (2) out-of-town suggestion satisfies the needs of users who travel to distant and unfamiliar areas. LBSNs can still suggest some POIs to the users but the results seem unreliable. It should be noticed that both the suggestion scenarios have its own spatio-temporal features.

In this study, we make the following original contributions:

(1) We model the multiple geo-social influence from social communities, time intervals and geographical locations to the textual content of reviews in both LBSNs and MBSNs based on the following facts: (a) users from a community might be interested in POIs located in the same region due to their frequent online communication; (b) the activity time of a community’s members tend to be tightly connected with each other; (c) a better POI suggestion result for a specific user can be inferred from review content extracted from LBSNs and MBSNs.

(2) We propose a latent probabilistic generative model called MTAS, which can accurately capture review words in LBSNs as well as MBSNs by taking into account information of social communities, time intervals and geographical locations. In addition, we can effectively infer the parameters of the MTAS model by the Gibbs sampling method.

(3) extensive experiments are conducted to evaluate the performance of the proposed MTAS model in two geo-social networks, and the experimental results show that our approach outperforms state-of-the-art baseline approaches in effectiveness and efficiency of POI suggestion.

The remainder of this paper is summarized as follows: Section 2 briefly reviews related works. Section 3 introduces some important definitions and formalize the POI suggestion problem. Section 4 describes the details of our proposed POI suggestion model, and then presents the inference process. Section 5 illustrates the application framework of POI suggestion. The experimental results are evaluated in Section 6. Lastly, we conclude this study in Section 7.