LSVP: A visual based deep neural direction learning model for point-of-interest recommendation on sparse check-in data
Introduction
With the popularity of location-based social networks (LBSNs) and mobile devices, users tend to share their locations and experiences, leading to a great amount of geo-tagged data to be accumulated. The availability of user check-in data brings in excellent opportunities to understand human behaviors and make recommendations to users.
Different from traditional recommendation, the check-in data for POI recommendation is extremely sparse. Because users must physically visit the POI to rate it, the cost is more expensive than a click or rating on a movie online. Besides, the cold start problem (no historical check-in records for new users or new POIs) is even more severe in personalized POI recommendation. Therefore, additional information needs to be incorporated to address these issues. Various information are integrated to traditional recommendation models, i.e., geographical influence [33], social correlations [13], [21] and textual contents [27], [25]. These information have been proven to be effective for improving POI recommendations.
We usually use users’ comments on POIs or the labels specified by the platform to obtain additional textual information [26]. However, in some social networks that dominate pictures such as Instagram and Flickr, users are willing to share photos on it than describe the attributes of POIs through text. To make a better recommendation for such users, we find that visual features in geo-tagged photos taken by users can provide meaningful context information for predicting users’ visit preferences [41]. For example, Fig. 1 shows some POI photos from two users. We can observe that user A prefers to visit the natural landscape while user B tends to visit the cultural place. From these photos, the POI information can be inferred, also users’ behaviors and preferences can be revealed. Previous work [18], [42] also suggests that photos and POIs have strong connections. Thus, incorporating photos could improve the performance of POI recommendations.
Unfortunately, most models that incorporate visual contents only capture users’ long-term preferences, which are considered statistical data or change slowly over time. Wang et al. [29] incorporate visual contents into a probabilistic model to learn the potential functions of users and POIs, Zhang et al. [39] integrates visual contents and geographic influence into a matrix decomposition model. However, the intention of user behavior is inherently variable. It will be affected by various factors during a specific period of time, such as the evolution of interests, instant demand, and global mainstream fashion [34]. Therefore, it is necessary to consider long-term, short-term and visual preferences of users.
In this paper, we integrate the visual contents of geo-tagged photos and the sequential patterns of check-in sequences into an attention-based neural model to capture users’ Long-term, Short-term, and Visual Preferences (LSVP). Specifically, we extract visual features from photos taken by users, and use them to understand the style of POIs and users’ visual preferences. Then, we adopt an LSTM-based network with temporal gating, which learns users’ long-term and short-term preferences from users’ historical check-ins. Finally, an adaptive attention mechanism is used to balance all the extracted users’ preferences. Experimental results show that our proposed model is significantly better than some of the latest models. The main contributions of this article are summarized as follows:
- •
We propose a deep neural network model LSVP to make POI recommendation for users who are active in social networks through pictures. We integrate visual contents and sequential patterns to achieve more accurate recommendation, so that the data sparsity problem can be well resolved.
- •
We design a two-layer attentional network for POI recommendation to effectively integrate the joint effects of users’ long-term, short-term and visual preferences through the specific contextual information, such that their weights are reasonably balanced for the personalized recommendation.
- •
We conduct extensive experiments on two real datasets. Experimental results demonstrate that our proposed model outperforms several state-of-the-art models significantly.
Section snippets
POI recommendation
In early works, there are some traditional recommendation methods based on Matrix Factorization (MF) and Collaborative Filtering (CF) [24], [43], which have been proven to be effective for lots of recommendation system. Due to the sparsity of check-ins and the large amount of POIs [1], [19], it is hard for traditional recommendation models to capture users’ preferences. To improve the accuracy of POI recommender system, some recent works [32] try to explore and integrate additional information
Problem definition
In this section, we first define the concepts used in this paper and then formalize the problem.
The overview of model
As shown in Fig. 2, we first crawl the photos from the public photo-sharing web site (i.e., Flickr). With the same approach described in [14], we obtain a list of POIs from Wikipedia and map these photos to user-POI visits. Furthermore, we construct users’ check-in sequences based on them. Second, We mainly model the users’ three aspects of preferences. Visual preferences are extracted from the user-generated photos by the VGG16 model [22], long-term preferences are extracted from the visited
Datasets
We apply the proposed model on the Yahoo! Flickr Creative Commons 100 M (YFCC100M) dataset [28], the largest public multimedia collection released, which consists of 100 million photos and 0.8 million videos posted on Flickr with relevant meta-information, such as the date taken, geo-location coordinates and geo-graphic accuracy. The geo-graphic accuracy ranges from the world level to the street level.
From this dataset, we use geo-tagged photos that were taken in Toronto and Budapest. More
Conclusion
In this paper, a hybrid model named LSVP is proposed for POI recommendation. LSVP models users’ short-term and long-term preferences from the check-in sequences, and it also learns users’ visual preferences through the geo-tagged photos. Then, these preferences are integrated into an attention network for personalized POI recommendation. We conduct experiments to evaluate the performance of our LSVP model on two real datasets. The results show the superiority of our proposal over other
CRediT authorship contribution statement
Yu Sang: Conceptualization, Methodology, Software, Writing - original draft. Huimin Sun: Methodology, Software, Writing - original draft. Chao Li: Data curation, Visualization, Validation. Lihua Yin: Methodology, Validation, Writing - review & 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.
Acknowledgements
This work is supported in part by the National Natural Science Foundation of China (No. 61872100) Researchand Development Program of China [Grant No. 2018AAA0100201,2018AAA010020].
Yu Sang received the B.S degree in business administration from the Huanghe S & T University, Zhengzhou, China, in 2015, and the M.S degree in social work from the Beijing City University, Beijing, China, in 2018. From 2018 to 2019, she worked as a Research Assistant in the Advanced Data Analytics Laboratory of Soochow University. Now she is also a Research Assistant in the Cyberspace Institute of Advanced Technology, Guangzhou University. Meanwhile, her current research interest is recommender
References (45)
- et al.
On workflow aware location-based service composition for personal trip planning
Fut. Gener. Comput. Syst.
(2019) - et al.
Vcg: exploiting visual contents and geographical influence for point-of-interest recommendation
Neurocomputing
(2019) - J. Chen, J. Xu, C. Liu, Z. Li, A. Liu, Z. Ding, Multi-objective spatial keyword query with semantics, in: DASFAA, Vol....
- et al.
r-tree: a pivot-based indexing structure for semantic-aware spatial keyword search
GeoInformatica
(2020) - C. Cheng, H. Yang, I. King, M.R. Lyu, Fused matrix factorization with geographical and social influence in...
- D.J. Crandall, L. Backstrom, D.P. Huttenlocher, J.M. Kleinberg, Mapping the world’s photos, in: WWW, 2009, pp....
- H. Gao, J. Tang, X. Hu, H. Liu, Exploring temporal effects for location recommendation on location-based social...
- J. Hays, A.A. Efros, IM2GPS: estimating geographic information from a single image, in: CVPR,...
- X. He, L. Liao, H. Zhang, L. Nie, X. Hu, T. Chua, Neural collaborative filtering, in: WWW, ACM, 2017, pp....
- et al.
Long short-term memory
Neural Comput.
(1997)
Recommending tours and places-of-interest based on user interests from geo-tagged photos
Recommending tours and places-of-interest based on user interests from geo-tagged photos
Where your photo is taken: geolocation prediction for social images
J. Assoc. Inf. Sci. Technol.
A context-aware personalized travel recommendation system based on geotagged social media data mining
Int. J. Geograph. Inf. Sci.
Semantic-aware top-k spatial keyword queries
WWW
Learning representations by back-propagating errors
Nature
Cited by (7)
Online content-based sequential recommendation considering multimodal contrastive representation and dynamic preferences
2024, Neural Computing and ApplicationsPoint of Interest recommendation for social network using the Internet of Things and deep reinforcement learning
2023, Mathematical Biosciences and Engineering
Yu Sang received the B.S degree in business administration from the Huanghe S & T University, Zhengzhou, China, in 2015, and the M.S degree in social work from the Beijing City University, Beijing, China, in 2018. From 2018 to 2019, she worked as a Research Assistant in the Advanced Data Analytics Laboratory of Soochow University. Now she is also a Research Assistant in the Cyberspace Institute of Advanced Technology, Guangzhou University. Meanwhile, her current research interest is recommender system.
Huimin Sun is a postgraduate student in Soochow University. Her research interests mainly on recommendation system.
Chao Li received the B.S and M.S degrees in computer science and technology from the Guilin University of Electronic Technology, Guilin, China, in 2007, and his Ph.D. degree in information security from University of Chinese Academy of Sciences, Beijing, China, in 2013. From 2013 to 2017, he was a Research Assistant in the Institute of Information Engineering, Chinese Academy of Sciences. He is currently an Associate Professor with the Cyberspace Institute of Advanced Technology, Guangzhou University. His current research interests include privacy preserving and access control in Internet of Things, and information leakage protecting.
Lihua Yin received her Ph.D. degree in computer science and technology from Harbin Institute of Technology, Harbin, in 2007. She is a professor at Cyberspace Institute of Advanced Technology, Guangzhou University. Her research interests include information security, bigdata privacy protection etc. She is a member of CCF and CIPS.