Elsevier

Applied Soft Computing

Volume 96, November 2020, 106536
Applied Soft Computing

Efficient point-of-interest recommendation with hierarchical attention mechanism

https://doi.org/10.1016/j.asoc.2020.106536Get rights and content

Highlights

  • In order to analyze more representative features in complex check-in data by POI recommendation methods based on machine learning, we define the concept of explicit features and implicit features. These concepts provide some ideas to collect data and select computational models for machine learning based methods.

  • In order to obtain more effective information from sparse data and to make use of high-contribution information as much as possible, we propose hierarchical attention mechanism with the structure of “local–global”. This mechanism focuses on the contribution degree of individual features to POI recommendation in a local way, and excavates the contribution degree and hidden information of POI recommendation from the combination features and overall features as a whole.

  • We first propose an NLP-based “User-POI” matching mechanism in the POI recommendation field. This mechanism mines “User-POI” matching degree1 from the semantic similarity of user comments and POI description. Then, we propose a fine-tuning function based on “User-POI” matching degree, which used “User-POI” matching degree to fine-tune the POIs predicted by the system to obtain more accurate POIs.

  • Considering the challenges of large-scale data operations, heavy user relationships maintenance, and cold start of new users, we constructed 3 data sets and several different evaluation scenarios to demonstrate the effectiveness of the hierarchical attention mechanism, “User-POI” matching degree, and the use of unstructured data. The experimental results show that the recommended performance of HAM-POIRec is optimal when compared with methods such as DeepPIM, SAE-NAD, MGMPFM and LRT, especially in predicting sequence POIs and cold-start problems.

Abstract

Personalized Point-of-Interest (POI) Recommendation is very important for application platforms based on Location Based Social Networks (LBSNs). It can assist users in making decisions to alleviate the problem of information overload, and can also improve the user experience of these platforms and advance platform operators achieve personalized and accurate advertising. However, there exist some problems of data sparseness and cold start for a single user, and it is also difficult to mine valuable long-tailed POIs, although the size of the check-in data is large. Therefore, in order to address the above problems, we propose a personalized POI Recommendation approach based on Hierarchical Attention Mechanism (HAM-POIRec) which can effectively increase data utilization. Firstly, we define the concepts of explicit features and implicit features, which pave the ideas of selecting data and computational models for POI recommendation based on machine learning. Secondly, we propose a hierarchical attention mechanism with the structure of local-to-global, which extracts contributions and mines more hidden information from individual features, combination features, and overall features. Finally, we present the Natural Language Processing (NLP)-based “User-POI” matching mechanism for the first time in the field of POI recommendation to improve the recommendation accuracy by fine-tuning the POIs predicted by the recommendation system. Extensive experiments are conducted for demonstrating that the HAM-POIRec method outperforms state-of-the-art DeepPIM method and the other comparison methods (SAE-NAD, MGMPFM and LRT), especially in predicting sequence POIs and solving cold start problem.

Introduction

With the rapid development of mobile devices and location acquisition technologies, it has become more and more convenient for people to obtain real-time location-based services [1], [2], [3]. In addition, the related platforms of Location Based Social Networks (LBSNs) including Yelp,2 Foursquare,3 Dianping4 and Mafengwo5 have also developed rapidly. On these platforms, the users share their locations and experiences in places they visit (such as tourist attractions, restaurants and shops, etc.), which generates a lot of check-in data [4], [5]. These places, the users like and have visited, are named as Point-of-Interest (POI) [6], [7]. A large number of interactive data between users and POIs bring potential data wealth to these mobile internet platforms based on LBSNs, but it causes the problem of information overload [8]. POI recommendation helps users to select their favorite POIs from the overloaded information. That is to say, on the one hand, the POI recommendation can understand the user’s personalized requirements to reduce the pressure of filtering information, and help the users understand their surrounding environment to assist them in making decisions. On the other hand, it can help platform operators to implement intelligent advertising services, and can improve the user experience of the platform while increasing its advertising revenue [9].

POI recommendation faces more challenges than general recommendation (such as product recommendation and movie recommendation) [10]. Firstly, users’ preferences for POIs are affected by geographic distance: Usually they visit a small number of POIs near their home or school. Second, the users may visit the same POIs (eg, workplace, breakfast shop, etc.) every day. Third, users’ preferences depend on time series, for example, the restaurants they visited in the early morning and late at night are different. Fourth, the users’ preferences are affected by social relationships, for example, he is affected by the preferences of his friends. Finally, data such as comments and descriptions of POIs can also influence user’s preferences. As shown in Table 1, in order to solve these challenges, POI recommendation has gradually formed a social network-based POI recommendation method and a POI recommendations based on check-in data context and topic content.

In the social network-based POI recommendation, collaborative filtering based on social relationships is a widely used method. This method implements recommendations based on the social relationships between two friends and the similarity of check-ins [11]. It does not consider all users, so the calculation speed is faster, but it also results in lower accuracy [12]. Immediately after, there appeared the Probabilistic Matrix Factorization with Social Regularization (PMFSR) for integrating social influences to learn the user’s potential preferences or potential characteristics using a matrix factorization method [13], [14]. However, none of these methods can handle large-scale data. Fortunately, the methods based on deep neural networks can solve this problem. For example, the self-attention autoencoders put forward a multi-dimensional attention mechanism to improve the accuracy of feature extraction by integrating geographic location, so that the model can obtain more hidden features to alleviate the problem of data sparsity [15]. However, this method still has the problem of being unable to mine high-quality long-tail POIs (that is, the cold start problem). Taking Foursquare and Twitter as an example, only about 20% of friends have checked in at the same POIs, which means that about 80% of friends do not have any POIs in common [16].

The POI recommendations based on check-in data context and topic content is the current mainstream method, which can alleviate the problem of sparse data and help the users to mine high-quality long-tail POIs [12], [17]. The core of this type of method is to extract more latent features from unstructured information such as rating, classification labels, and description text to describe users and POIs, and build their associations based on these latent features [18], [19], [20]. Previously, referring to the user-based and item-based collaborative filtering on the use of ratings to complete recommendations, researchers have used ranking-based geographic factorization machine to complement their potential preferences based on geographic location and check-in context data to recommend unvisited POIs for users [13], and use a collaborative POI recommendation based on naive bayes to model user check-in behavior using power law distribution to predict user access probability to new POIs [21]. However, the above methods usually model user preferences based on the weighted average of the ratings or the inner product of latent factors. It is the calculation of the inner product that limits the expressive power of the POI recommendation, which makes the method unable to process large-scale data [12], [22], [23]. As shown in Table 1, researchers have attempted to extract more latent features from structured and unstructured data by using basic Recurrent Neural Network (RNN), and achieved good results, but with problems such as gradient disappearance [24]. Hence, the researchers put forward two solutions: one is to introduce attention mechanism to extract more useful expression features from limited data [25], [26], the other is to add unstructured data to mine more valuable potential features [18], [19], [20], [25] [26]. However, these methods still ignore the latent features of the combination features and overall features (including the different contributions of these features to the recommendation). In addition, no researchers has yet attempted to extract the “User-POI” matching degree from text similarity to improve recommendation performance.

For the sake of solve the above problems, extracting more expression features from the sparse data of individual users to complete accurate personalized POI recommendation, and mine high-quality long-tail POIs for users, we propose a POI Recommendation method based on Hierarchical Attention Mechanism termed HAM-POIRec. First of all, in addition to the review text shown in Table 1, we have also added other unstructured data including POI description text and pictures, to further refine the concept of explicit features extracted from structured features and implicit features from unstructured data. Secondly, in the input data of the neural network, there are differences in the contribution of individual features, combination features and overall features to the recommendation, and it contains a lot of useful hidden information. Therefore, we propose a hierarchical attention mechanism of the local-to-global structure to extract more hidden information for sparse data and improve the utilization of high-quality information. Finally, we extract the potential association of “User-POI” from user comment (text) and POI description (text) through Natural Language Processing (NLP), and use this association as weights to fine-tune predictors built from the users and POIs features to improve recommendation performance. Overall, the major contributions of this paper are as follows:

  • In order to analyze more representative features in complex check-in data by POI recommendation methods based on machine learning, we define the concept of explicit features and implicit features. These concepts provide some ideas to collect data and select computational models for machine learning based methods.

  • In order to obtain more effective information from sparse data and to make use of high-contribution information as much as possible, we propose hierarchical attention mechanism with the structure of “local–global”. This mechanism focuses on the contribution degree of individual features to POI recommendation in a local way, and excavates the contribution degree and hidden information of POI recommendation from the combination features and overall features as a whole.

  • We first propose an NLP-based “User-POI” matching mechanism in the POI recommendation field. This mechanism mines “User-POI” matching degree6 from the semantic similarity of user comments and POI description. Then, we propose a fine-tuning function based on “User-POI” matching degree, which used “User-POI” matching degree to fine-tune the POIs predicted by the system to obtain more accurate POIs.

  • Considering the challenges of large-scale data operations, heavy user relationships maintenance, and cold start of new users, we constructed 3 data sets and several different evaluation scenarios to demonstrate the effectiveness of the hierarchical attention mechanism, “User-POI” matching degree, and the use of unstructured data. The experimental results show that the recommended performance of HAM-POIRec is optimal when compared with methods such as DeepPIM, SAE-NAD, MGMPFM and LRT, especially in predicting sequence POIs and cold-start problems.

The organization of this paper is as follows: Section 2 briefly reviews the literature on POI recommendation system. Section 3 describes the problem formulation. Section 4 describes the framework and implementation details of the HAM-POIRec. Section 5 reports the results of experiments followed by the conclusions and future work in Section 6.

Section snippets

Related works

In the past few years, many researchers put forward the solution for recommending new locations to the users into practice, which has gradually evolved into the research field of POI recommendation [27]. The POI recommendation methods mainly include two categories: One is a POI recommendation-based on Social network; the other is a recommendation method based on check-in data context and topic content, which has recently started to use the deep learning methods based on unstructured data.

Problem formulation

One of the most difficult challenges for mobile internet platforms based on LBSNs is how to accurately predict the next POI. As shown in Fig. 1, in the process of solving this challenge, the researches are using social network methods that use information such as relationship, time and distance have become common, and the deep learning method based on text information has emerged. Although this information contains many hidden features that improve the accuracy of POIs prediction, we observe

Proposed approach

In order to solve the problems as shown in Eq. (1), we propose a HAM-POIRec method based on the hierarchical attention mechanism. For the sake of obtaining more accurate auxiliary information from structured data and unstructured data, this method proposes a hierarchical attention mechanism to obtain more latent features, and use text similarity to strengthen and improve the accuracy of recommendations. For example, for a user who often works overtime late at night, the model can analyze the

Datasets

We utilize the Yelp Dataset Challenge Round 137 of the world’s largest review site. For the sake of verifying the comprehensive performance of the algorithm, as shown in Table 2, we construct a large-scale data set yelp-b, a cold-start user data set yelp-c, and a heavy user data set yelp-h (the heavy user refers to the old users who reuse a product or service) according to different screening conditions. Among them, Nt represents the length of the string, Nc is the

Conclusions

In this paper, we propose a personalized POI recommendation based on the hierarchical attention mechanism. Firstly, the approach integrates structured and unstructured data from Yelp, and on this basis we refined the concepts of explicit features and implicit features to guide the model to design different computing components based on different data. Then, we propose a hierarchical attention mechanism that extracts contributions and more hidden information from individual features, combination

CRediT authorship contribution statement

Guangyao Pang: Conceptualization, Methodology, Software, Validation, Investigation. Xiaoming Wang: Resources, Supervision, Project administration, Funding acquisition. Fei Hao: Formal analysis, Writing - original draft. Liang Wang: Formal analysis, Writing - original draft. Xinyan Wang: Writing - original draft.

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.

Acknowledgments

This work was supported by the National Natural Science Foundation of China (Grant Nos. 61872228, 61702317, 61562074, 61862056), the National Key R&D Program of China (Grant No. 2017YFB1402102), the Shaanxi Provincial Key R&D Plan of China (Grant No. 2020ZDLGY10-05), the Natural Science Basic Research Plan in Shaanxi Province of China (Grant No. 2019JM-379, 2018JQ6048), the Guangxi Science and Technology Planning Project, China (Guike AB16380273), the Guangxi Innovation-Driven Development

References (51)

  • GaoR. et al.

    Exploiting geo-social correlations to improve pairwise ranking for point-of-interest recommendation

    China Commun.

    (2018)
  • HaoF. et al.

    Mobile cloud services recommendation: a soft set-based approach

    J. Ambient Intell. Humaniz. Comput.

    (2018)
  • YuZ. et al.

    Personalized travel package with multi-point-of-interest recommendation based on crowdsourced user footprints

    IEEE Trans. Hum.-Mach. Syst.

    (2016)
  • HeJ. et al.

    Inferring a personalized next point-of-interest recommendation model with latent behavior patterns

  • YuY. et al.

    A survey of point-of-interest recommendation in location-based social networks

  • MaH. et al.

    Recommender systems with social regularization

  • LiuQ. et al.

    Predicting the next location: A recurrent model with spatial and temporal contexts

  • ChengC. et al.

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

  • LianD. et al.

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

  • MaC. et al.

    Point-of-interest recommendation: Exploiting self-attentive autoencoders with neighbor-aware influence

  • H. Gao, J. Tang, H. Liu, gSCorr: modeling geo-social correlations for new check-ins on location-based social networks,...
  • LiX. et al.

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

  • PangG. et al.

    ACNN-FM: A novel recommender with attention-based convolutional neural network and factorization machines

    Knowl.-Based Syst.

    (2019)
  • ZhangD.Y. et al.

    Large-scale point-of-interest category prediction using natural language processing models

  • Q. Yuan, G. Cong, A. Sun, Graph-based point-of-interest recommendation with geographical and temporal influences, in:...
  • Cited by (0)

    1

    Matching degree indicates the possibility that the user likes the POI..

    View full text