Personalized location recommendation by fusing sentimental and spatial context
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 be the set of users and be the set of POIs, where M and N denote the number of users and POIs respectively. is a user-POI matrix with 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)
- et al.
CAPER: Context-aware personalized emoji recommendation
IEEE Trans. Knowl. Data Eng.
(2020) - et al.
An adaptive point-of-interest recommendation method for location-based social networks based on user activity and spatial features
Knowl.-Based Syst.
(2019) - et al.
Viewpoint recommendation based on object oriented 3d scene reconstruction
IEEE Trans. Multimedia
(2020) - et al.
Word sense disambiguation: A comprehensive knowledge exploitation framework
Knowl. Based Syst.
(2020) - et al.
Recommendations in location-based social networks: a survey
GeoInformatica
(2015) - H. Li, Y. Ge, D. Lian, H. Liu, Learning user’s intrinsic and extrinsic interests for point-of-interest recommendation:...
- H. Wang, H. Shen, W. Ouyang, X. Cheng, Exploiting POI-specific geographical influence for point-of-interest...
- Z. Yao, Exploiting human mobility patterns for point-of-interest recommendation, in: Proc. WSDM, 2018, pp....
- et al.
Learning travel recommendations from user-generated GPS traces
ACM TIST
(2011) - E. Cho, S.A. Myers, J. Leskovec, Friendship and mobility: user movement in location-based social networks, in: Proc....
Author topic model-based collaborative filtering for personalized POI recommendations
IEEE Trans. Multimedia
Are they different? Affect, feeling, emotion, sentiment, and opinion detection in text
IEEE Trans. Affect. Comput.
Rating prediction based on social sentiment from textual reviews
IEEE Trans. Multimedia
Inferring social ties between users with human location history
J. Ambient Intell. Humaniz. Comput.
Exploring users’ internal influence from reviews for social recommendation
IEEE Trans. Multimedia
Service quality evaluation by exploring social users’ contextual information
IEEE Trans. Knowl. Data Eng.
Personalized reason generation for explainable song recommendation
ACM TIST
SeCo-LDA: Mining service co-occurrence topics for composition recommendation
IEEE Trans. Serv. Comput.
Augmented collaborative filtering for sparseness reduction in personalized POI recommendation
ACM TIST
Location recommendation for enterprises by multi-source urban big data analysis
IEEE Trans. Serv. Comput.
Cited by (60)
Neuro-symbolic recommendation model based on logic query
2024, Knowledge-Based SystemsImproving the spatial–temporal aware attention network with dynamic trajectory graph learning for next Point-Of-Interest recommendation
2023, Information Processing and ManagementDynamic context management in context-aware recommender systems
2023, Computers and Electrical EngineeringAspect-based sentiment analysis via multitask learning for online reviews
2023, Knowledge-Based SystemsTeSP-TMF: A temporal-aware personalized POI recommendation approach based on potential preferences and grey relational analysis
2023, Electronic Commerce Research and ApplicationsRanking-based contrastive loss for recommendation systems
2023, Knowledge-Based SystemsCitation 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].