Personalized location recommendation by fusing sentimental and spatial context

https://doi.org/10.1016/j.knosys.2020.105849Get rights and content

Abstract

Internet users would like to obtain interesting location information for a travel. With the rapid development of social media, many kinds of location recommender systems are proposed in recent years. Existing methods mostly focus on mining user check-in information that could be leveraged to understand their trajectories. However, the characteristics and attributes of geographical locations also play an important role in recommender systems. In this paper, sentimental attributes of locations are explored and we propose a Point of Interest (POI) mining method and a personalized recommendation model by fusing sentimental spatial context. First, a Sentimental–Spatial POI Mining (SPM) method is utilized to mine the POIs by fusing the sentimental and geographical attributes of locations. Second, we recommend the POIs to users by a Sentimental–Spatial POI Recommendation (SPR) model incorporating the factors of sentiment similarity and geographical distance. Last, the advantages and superior performance of our methods are demonstrated by extensive experiments on a real-world dataset.

Introduction

In recent years, social networks have a significant development. Through location based social networks (LBSNs) on mobile devices or online, users can share their geographical position information and check-ins. Social network services also encourage them to share their experiences, reviews, ratings, photos, and moods. Such information brings new opportunities for recommender systems. Existing location recommender systems mostly focus on exploring user information [1], which includes users’ profiles [2], locations [3], and trajectories [4], [5], [6]. Features of locations also attract researchers, such as the frequency of visiting by users and the category attribute of locations [1]. However, in the process of POI mining, the sentimental features of locations are seldom considered. The recommendations may not suit users’ sentimental preference. For example, a user living in downtown wants to enjoy a natural and peaceful place to have a rest, but recommender systems mainly focus on the geographical attributes of nearby locations, and it may recommend some POIs full of people and noise. It cannot meet the user’s needs. It indicates that POIs have not only geographical attributes but also sentimental attributes, which is an major factor that should be exploited. The POIs with many historical spots are solemn, and the POIs which have business districts and clubs are lively, whereas the POIs with a lot of trees and pools are peaceful and relaxing. The sentimental attributes of POIs can be discovered by analyzing the data on social networks.

Users share their experiences and locations on the websites such as Twitter1 and Sina Weibo2 by checking-in. As shown in Fig. 1, Text is the user’s comment about his/her status or feelings. Location shows the GPS position. The information of Text is strongly connected with the Location in some way. In this study, through sentiment analysis of Text, the sentimental attributes of the Location could be discovered. After that, POIs with obvious sentimental attributes are mined and will be recommended if they are nearby and matching users’ preferences.

First, the Sentimental–Spatial POI Mining (SPM) method is proposed to mine the POIs with obvious sentimental attributes. According to the result of sentiment analysis and GPS positions, the POIs which have a thick density of social media data and similar sentimental attributes are discovered. Second, they are recommended to different users by our Sentimental–Spatial POI Recommendation (SPR) model. It incorporates the factors of sentiment similarity and geographical distance. It is based on the widely adopted latent factor model realized by Probabilistic Matrix Factorization (PMF) [7]. A POI with a higher sentiment score should be ranked above a POI with a lower sentiment score. However, users have different preferences for topics [8]. For example, for a restaurant, some users think the price is too high, but others may prefer the unique taste and do not care its price. Thus, the POI with a higher sentiment score does not mean it must be better than a POI with a lower sentiment score for a particular user. Therefore, we prefer to leverage sentiment similarity rather than use absolute sentiment score to optimize the latent features of POIs in our model.

Note that, affect is non-conscious and it is a fundamental and broader concept. Feelings and emotions are the conscious expressions of affect, while the sentiment is a high-level conscious attitude, also is an emotional disposition [9]. In this paper, the topic-based sentiment [10] is a more accurate and straight term than affective attributes to represent user preference and POI attributes.

Our contributions are shown as follows:

  • We propose a POI Mining method and a personalized POI Recommendation method by fusing sentimental and spatial context. We explore the rich textual descriptions and users’ geographical information and propose new features and factors in our methods. Experiment results demonstrate the superior performance of our methods.

  • In our POI Mining method, the sentiment context is proposed as a new attribute of the POI. Additionally, we remove the redundant and noisy microblog posts by a temporal filter to improve the mining performance. Through the POI Mining method, we could discover the POIs with salient sentimental attributes.

  • In our POI Recommendation method, we propose two new factors: the factor of sentiment similarity between POIs and the factor of geographical distance between user’s multi-activity centers and POIs. We incorporate both of them into Probabilistic Matrix Factorization model for POI recommendation.

The main differences between this paper and our previous work [11] are: (1) we improve SPM method by incorporating spatial and temporal information to remove the redundancy and noise of the data; (2) we improve SPR model by exploring the user’s multi-activity centers; (3) we improve the readability of our model by summarizing a whole procedure of our algorithm, and present more details of model training; (4) more datasets, experiments and discussions are implemented.

The rest of this paper is organized as follows. We review the related work on recommender systems and the sentiment analysis in Section 2. The details of our SPM method and SPR model are presented in Section 3 and Section 4. Section 5 gives experiment results and discussions, and Section 6 concludes this paper.

Section snippets

Related work

In this section, we first introduce some related work on recommender systems in LBSNs and POI recommendation, and then some methods of sentiment analysis are reviewed. Additionally, we also discuss the main differences between our work with related work.

Our sentimental–spatial POI mining method

Fig. 2 shows the overview of our SPM method which incorporates POIs’ geographical attributes and sentimental attributes. In Fig. 2, the red, green, and blue dot represents the microblog posts with high, fair, and low sentiments respectively. There are two steps in SPM method. First, the sentimental features of locations are discovered by calculating the sentiment of posts in these locations. Second, based on sentiment values and GPS positions of these microblog posts, the locations which have a

Our sentimental–spatial POI recommendation model

In this section, the SPR model will be introduced in details. This recommendation model utilizes low-rank probabilistic matrix factorization [7] to figure out how much a user prefers a POI. Let U={u1,u2,,uM} be the set of users and P={i1,i2,,iN} be the set of POIs, where M and N denote the number of users and POIs respectively. R=[Ru,i]M×N is a user-POI matrix with Ru,i representing the number of check-ins made by user u at POI i. In this paper, a user’s preference on a POI is represented by

Experiment

Adopting the data of Sina Weibo as the dataset, the proposed models are verified by experiments. We compare the performance of our SPM method with Meanshift method [53], K-means, and Gaussian Mixture Model (GMM). They are traditional popular clustering methods. Then our SPR model is compared with BaseMF [7], Bias-MF [54], CF-MF [49], LBSMF [35], and IRenMF [32], DeepMF [47] and CSPR [11]. In addition, we present POIs mined by SPM on the map.

Conclusion

Existing location recommender systems mainly focus on physical characteristics of locations rather than the sentimental attributes of locations. The SPM and SPR methods are proposed in the paper. According to our methods, the POIs with obvious sentimental attributes are mined and recommended to users. We conducted extensive experiments on a large real-world dataset and demonstrated that the proposed methods have better effectiveness than existing approaches.

In our future work, check-in

CRediT authorship contribution statement

Guoshuai Zhao: Conceptualization, Methodology, Visualization, Writing - original draft, Formal analysis. Peiliang Lou: Methodology, Software, Data curation, Writing - original draft, Formal analysis. Xueming Qian: Resources, Supervision. Xingsong Hou: Writing - reviewing & editing.

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 in part by the NSFC, China under Grants 61902309, 61732008, 61772407, and 1531141; in part by the World-Class Universities (Disciplines) and the Characteristic Development Guidance Funds for the Central Universities, China (PY3A022); in part by the Fundamental Research Funds for the Central Universities, China; in part by China Postdoctoral Science Foundation; and in part by the National Postdoctoral Innovative Talents Support Program, China for G. Zhao.

References (58)

  • R. Salakhutdinov, A. Mnih, Probabilistic matrix factorization, in: Proc. NIPS, 2007, pp....
  • JiangS. et al.

    Author topic model-based collaborative filtering for personalized POI recommendations

    IEEE Trans. Multimedia

    (2015)
  • MunezeroM. et al.

    Are they different? Affect, feeling, emotion, sentiment, and opinion detection in text

    IEEE Trans. Affect. Comput.

    (2014)
  • LeiX. et al.

    Rating prediction based on social sentiment from textual reviews

    IEEE Trans. Multimedia

    (2016)
  • P. Lou, G. Zhao, X. Qian, H. Wang, X. Hou, Schedule a rich sentimental travel via sentimental POI mining and...
  • XiaoX. et al.

    Inferring social ties between users with human location history

    J. Ambient Intell. Humaniz. Comput.

    (2014)
  • M. Ye, D. Shou, W. Lee, P. Yin, K. Janowicz, On the semantic annotation of places in location-based social networks,...
  • J. Zhang, C. Chow, iGSLR: personalized geo-social location recommendation: a kernel density estimation approach, in:...
  • Y. Zheng, L. Zhang, X. Xie, W. Ma, Mining interesting locations and travel sequences from GPS trajectories, in: Proc....
  • R. Raymond, T. Sugiura, K. Tsubouchi, Location recommendation based on location history and spatio-temporal...
  • ZhaoG. et al.

    Exploring users’ internal influence from reviews for social recommendation

    IEEE Trans. Multimedia

    (2019)
  • ZhaoG. et al.

    Service quality evaluation by exploring social users’ contextual information

    IEEE Trans. Knowl. Data Eng.

    (2016)
  • J. Sang, T. Mei, J. Sun, C. Xu, S. Li, Probabilistic sequential POIs recommendation via check-in data, in: Proc. ACM...
  • L. Wei, Y. Zheng, W. Peng, Constructing popular routes from uncertain trajectories, in: Proc. ACM SIGKDD, 2012, pp....
  • ZhaoG. et al.

    Personalized reason generation for explainable song recommendation

    ACM TIST

    (2019)
  • GaoZ. et al.

    SeCo-LDA: Mining service co-occurrence topics for composition recommendation

    IEEE Trans. Serv. Comput.

    (2019)
  • CuiC. et al.

    Augmented collaborative filtering for sparseness reduction in personalized POI recommendation

    ACM TIST

    (2017)
  • S. Wang, Y. Wang, J. Tang, K. Shu, S. Ranganath, H. Liu, What your images reveal: Exploiting visual contents for...
  • ZhaoG. et al.

    Location recommendation for enterprises by multi-source urban big data analysis

    IEEE Trans. Serv. Comput.

    (2017)
  • Cited by (60)

    • Dynamic context management in context-aware recommender systems

      2023, Computers and Electrical Engineering
    • Ranking-based contrastive loss for recommendation systems

      2023, Knowledge-Based Systems
      Citation Excerpt :

      They have become a fundamental technology and an important research area in the big data era. Recommendation algorithms have evolved from traditional methods to deep learning-based methods [1–3], as well as the popular graph convolution network (GCN) [4–6]. Top-k recommendation is one of the basic tasks in recommendation systems [7,8].

    View all citing articles on Scopus
    View full text