Location recommendation by combining geographical, categorical, and social preferences with location popularity

Highlights

• A unified location recommendation framework is proposed to measure users’ check-in probability of unvisited locations.

• A novel method based on category hierarchy and semantic similarity between location tags is proposed to model categorical preference.

• Comprehensive experiments are conducted to show that our method outperforms other state-of-the-art baselines.

• Location popularity, exploited as a global attribute of locations, can result in a significant improvement on recommendation performance.

• More friends do little help in improving the precision.

Abstract

The primary aim of location recommendation is to predict users’ future movement by modeling user preference. Multiple types of information have been adopted in profiling users; however, simultaneously combining them for a better recommendation is challenging. In this study, a novel location recommendation method that incorporates geographical, categorical, and social preferences with location popularity is proposed. Experimental results on two public datasets show that the proposed method significantly outperforms two state-of-the-art recommendation methods. Geographical preference generally shows more importance than both categorical and social preferences. A category hierarchy that unleashes the independent assumption of location tags improves categorical preference. Location popularity proves to be a useful metric in ranking candidate locations. The findings of this study can provide practical guidelines for location recommendation services.

Introduction

A large number of location-based social networks (LBSNs), such as Yelp and Foursquare, have attracted millions of users to share points of interest (POIs, i.e., locations) and social friendship via check-ins (Cheng, Yang, King & Lyu, 2016). Users would check in at locations such as restaurants, bars, home service stores, as well as museums, art galleries, and parks to comment about the services or share their experiences. By mining these data, user behaviors can be widely applied to recommend locations according to user-derived preferences (Bao, Zheng, Wilkie & Mokbel, 2015). As a crucial feature of LBSNs, location recommendation has received extensive attention from industry and academia. By mining users’ check-in records and social relationships, location recommendation is useful for suggesting new places for users to visit. It can be defined as recommending an unvisited location list to a user based on his/her check-in sequences, including visited locations and check-in time stamp (Zhao, King & Lyu, 2016). This function can not only increase the visibility of locations and attract more customers, but also improve user experience and customer satisfaction (Wu, Kao, Wu & Huang, 2015; Zhu, Chang, Luo & Li, 2014). One classic application is to free users from information overload dilemma and help them discover unvisited locations that satisfy their personalized preferences.

Location recommendation aims to predict users’ future movement by mining their check-in histories or consuming behaviors. In this process, massive volume of location data and high sparsity of user-visit data render significant challenges for improving the recommendation quality. For example, the Yelp dataset¹ used in this study involves 393 cities with an average of 353 locations. Las Vegas owns the most venues, i.e., 28,670 in total. However, the average number of user check-ins is only 3.75, involving 1.48 cities. As shown in Fig. 1, more than half of the users (53%) only have one check-in record, and those who have less than five check-ins constituted 86.75% of the users. Most users (73.43%) have visited only one city, while only 1.44% of the users have visited more than five cities. Data sparsity limits the accuracy in inferring personalized user preferences (Liu, Wu & Liu, 2013), rendering it difficult to discover potential locations of interest from massive venues. Therefore, the crucial problem to be solved by location recommendation is to precisely model preferences from exiguous visit records.

Recent studies have applied collaborative filtering (CF), content-based methods, and hybrid methods to extract useful information from multiple types of data for modeling user preferences in location recommendation (Rehman, Khalid & Madani, 2016). CF infers the similarities between users and venues by identifying similar visiting patterns for users and venues from user behaviors. Some important properties, such as business category and geographical characteristics, are not considered in CF but are exploited in content-based methods. In modeling users’ preferences, the importance of geographical, categorical, and social information have been demonstrated (Zhao et al., 2016). Geographical information defines the physical range of user activity, categorical information reflects the types of venues favored by users, while social information reveals the social influence of locations on users. Simultaneously combining these three aspects may help alleviate problems caused by sparse data; however, this has been proven challenging (Geng, Jiao, Gong, Li & Wu, 2019). Early studies only adopted one or two aspects, while a few recent methods attempted to exploit all three aspects to model user preferences from a broader perspective (Ference, Ye & Lee, 2013; Leung, Lee & Lee, 2011; Noulas, Scellato, Lathia & Mascolo, 2012). In addition to the three aspects above, word-of-mouth opinions measured as location popularity from the public can influence users’ decision making (Liu, Fu, Yao & Xiong, 2013).

In this study, we propose a new location recommendation approach that measures users’ personalized preferences to locations by geographical, categorical, and social associations, as well as the popularity of locations. Although neither the three aspects of preferences nor location popularity are novel, we introduce a new method to integrate them into the model to improve the accuracy of location recommendation. Specifically, the main contributions of this study can be summarized as follows:

• A unified location recommendation framework that combines geographical, categorical, and social preferences with location popularity is proposed to measure users’ check-in probability of unvisited locations.

• A novel method is developed based on category hierarchy to capture categorical preference by calculating the semantic similarity between location tags.

• Unlike existing studies, location popularity is exploited as a global attribute of location in this study. The priority ranks of locations derived by all three types of preferences are further adjusted by location popularity.

• The proposed recommendation approach is experimentally evaluated on two large-scale datasets collected from Weeplaces and Yelp; subsequently, it is compared with two state-of-the-art location recommendation baselines.

The remainder of this paper is organized as follows. Related studies are reviewed in Section 2. The proposed method is introduced in Section 3 and compared with two state-of-the-art baselines on two datasets in Section 4. Subsequently, experimental results are discussed in Section 5, followed by conclusions and future studies in Section 6.