A personalized point-of-interest recommendation model via fusion of geo-social information

Abstract

Recently, as location-based social networks (LBSNs) rapidly grow, general users utilize point-of-interest recommender systems to discover attractive locations. Most existing POI recommendation algorithms always employ the check-in data and rich contextual information (e.g., geographical information and users’ social network information) of users to learn their preference on POIs. Unfortunately, these studies generally suffer from two major limitations: (1) when modeling geographical influence, users’ personalized behavior differences are ignored; (2) when modeling the users’ social influence, the implicit social influence is seldom exploited. In this paper, we propose a novel POI recommendation approach called GeoEISo. GeoEISo achieves three key goals in this work. (1) We develop a kernel estimation method with a self-adaptive kernel bandwidth to model the geographical influence between POIs. (2) We use the Gaussian radial basis kernel function based support vector regression (SVR) model to predict explicit trust values between users, and then devise a novel trust-based recommendation model to simultaneously incorporate both the explicit and implicit social trust information into the process of POI recommendation. (3) We develop a unified geo-social framework which combines users’ preference on a POI with the geographical influence as well as social correlations. Experimental results on two real-world datasets collected from Foursquare show that GeoEISo provides significantly superior performances compared to other state-of-the-art POI recommendation models.

Introduction

Nowadays, location-based social networks (LBSNs), such as Foursquare and Gowalla, has been growing rapidly, due to the advances in location-acquisition and wireless communication technologies. In LBSNs, users check in some interesting locations and share their travel experiences by uploading photos, providing ratings and comments. Since LBSNs have collected a huge volume of users’ activity records, LBSNs have become an ideal platform to investigate users’ online behavior patterns. Many interesting and valuable applications can be built on this platform, such as location recommender system [1], [2], traveling routes recommendation [3], [4], extracting robust feature of location from remote sensing images [5], [6], and so on. As an important component of LBSNs, POI recommendation first helps users to explore new POIs. Then, it also helps advertising agencies to launch location-aware personalized services for potential customers and improve business profits [7].

Most existing POI recommendation methods [8], [9], [10] apply the collaborative filtering techniques with the user-POI check-in matrix to compute the preference score between a user and an unvisited POI. However, the user-POI check-in matrix is highly sparse with numerous missing entries, because users have only visited a very small proportion of POIs in an LBSN. As a result, these methods usually suffer from low recommendation quality. Since the Netflix-prize competition, the recommendation algorithm based on matrix-factorization (MF) techniques [11] have received extensive attention from academia and industry due to their good scalability and accurate predictive ability to deal with large-scale data. Several sophisticated and highly efficient MF-based models have been proposed. Representative ones of this kind include SVD++ model [12], probabilistic MF model [13], and nonparametric MF model [14].

In recent years, further research on matrix factorization yields more sophisticated models, e.g., LF-based CF model via second-order optimization model [15], weighted trace-norm regularization-based model [16], collaborative Gaussian process-based preference model [17], alternating direction method-based nonnegative latent factor model [18], [19], [20], [21], [22]. Meanwhile, these ideas are also used to address other relative issues. e.g., QoS prediction [23], [24], heavy computational overhead in ELM [25], particle swarm optimization [26], [27], [28], kalman filter extension [29], [30], video re-indexing [31], and mobile-user tracking [32].

Compared with the conventional recommendation systems, POI recommender systems using sparse geo-social networking data are more complex in LBSNs, which have the following unique features:

(1) Geographic influence: in LBSN, POIs are encoded with latitude and longitude, which distinguishes POIs form other items, such as books, music and movies in conventional recommender systems, which is called geographical influence.

(2) Social influence: in LBSNs, a person may prefer POIs highly recommended by her friends. In other word, social friends are tend to share more common interest than strangers, which is called social influence.

Based on the unique characteristics of the POI recommendation system mentioned above, some studies [33], [34] have turn to integrate the geographical influence with users’ social influence in LBSNs to enhance the quality of location recommendations. The social friendships is exploited as a user-user similarity matrix to improve the quality of location recommendation. Unfortunately, these studies suffer from two major limitations.

(1) The majority of existing POI recommendation algorithms have come up with a variety of geographic information modeling systems [35], [36], [37], [38] in light of the First Law of geography. For example, The literatures [35], [36] assume that there is an inverse relationship between a user’ s preference to a POI and the distance from a certain POI to the POI that the user has visited. The literatures [37] assume that the distance of visited locations follows a power-law distribution (PD), while in other document [38] a multi-center Gaussian model is proposed on the basis of the gathering effect in terms of the locations of users’ check-in. In these studies, the geographical influence of locations is universally modeled as a common distance distribution for all users or applied the same inversely proportional relation between the propensity and the distance for all users. But multiple factors (e.g., other users’ past check-in activities and the user’ s own personality etc.) can reveal the real geographical influence of locations on users’ check-in behaviors. For instance, indoorsy persons are inclined to conduct check-in POIs around their houses, while outdoorsy persons are more likely to travel around the world and explore new POIs. We focus on real-world examples who are randomly chosen from Foursquare [39] to show that a user’ s check-in behavior is unique. In Fig. 1, we can observe that the distribution of check-ins of the three users is quite different and extremely personalized.

(2) Currently, in some references [40], [41], [42], [43], the similarities between users are seamlessly connected to the collaborative filtering recommendation models. In general, when computing users’ similarities, such similarity measurement only considers those items or users sharing similar ratings for the purpose of easy calculation, which makes the predicting accuracy of the similarity measurement remains at a low level. Some researchers at present have integrated the users’ degree of trust as one of the most important types of social information into the recommendation systems when trying to improve the recommendation quality and have made certain achievements so far [44], [45], [46], [47], [48], [49]. However, the improvement of most trust social information-based recommender systems still have its inherent limitations, and the explanations are as follows: (1) in LBSNs, the fact is that a user only trusts a few other users, leaving a large number of fuzzy and sparse relationships among users. In other words, the data collected from LBSNs is imbalanced, and the explicit trust relationship between users form a non-linear relationship; (2) given very sparse trust networks exist, most existing works only consider users’ explicit trust relationship value, and ignore many valuable implicit trust relations as well as the characteristics; (3) friends with mutual trust are impossible to have completely similar interests. Therefore, explicit trust relationship may involves too much noise, which consequently impacts the performance of recommender systems.

In this paper, we propose a novel personalized geo-social POI recommendation model, called GeoEISo. The contributions of our work are summarized below:

(1) To fully consider users’ personalized check-in behaviors to model geographic influence, we propose to adopt a kernel function density estimation method with an adaptive bandwidth in different level to model geographic influence, which can obtain different distribution forms according to users’ check-in activities. Actually, to best our knowledge, no previous works investigate the personalized geographical influence of locations on a user’ s check-in behavior through learning different kernel bandwidths when adopting kernel density estimation algorithms.

(2) Incorporating trust social information into recommender systems has demonstrated potential to improve social recommendation performance, we propose a novel trust-based social recommendation model based on SVD++ model, which both the explicit and implicit trust relationship of user and ratings are involved to generate predictions. In addition, to the authors’ knowledge, this is the first work that adopt support vector regression based on the Gaussian radial basis function to model the explicit trust relationship to make sure outstanding learning ability and satisfactory precision, indicating its novelty.

(3) To improve the quality of POI recommendations, we design of decomposition model based on the matrix factorization model in an effort to integrate the social information and geographical information.

(4) We conduct extensive experiments to evaluate the performance of GeoEISo on two large-scale real-world data sets. Experimental results show that GeoEISo outperforms other state-of-the-art POI recommendation algorithms.

The rest of this paper is organized as follows. Section 2 briefly reviews related work in POI recommender systems. Section 3 introduces the preliminary knowledge used in this paper. Section 4 describes the details of our proposed POI recommendation algorithm. Experiments are evaluated in Section 5. Finally, we conclude this paper and present some directions for future work in Section 6.