Towards real-time demand-aware sequential POI recommendation
Introduction
Point-of-Interest (POI) recommendation has attracted increasing attention in recent years due to the proliferation of location-based social networks (LBSNs) such as Gowalla and Foursquare. Millions of users share their experience and opinions of visiting POIs like restaurants and stores every day, making it possible to predict where users like to visit next. During the last decade, many POI recommendation algorithms have been proposed to recommend users with a list of POIs according to their preference (e.g., to recommend an art lover to visit museums). To learn such user’s preferences, several representative studies have been introduced such as collaborative filtering (CF) models [1], [2] and artificial neural models [3], [4]. However, the temporal visiting patterns are little considered in these traditional methods.
Compared to general POI recommendation, next POI recommendation aims to predict the next POI based on the historical check-in activity sequence of a user. Following this line, numerous next POI recommendation algorithms have been proposed, and most of them try to learn user’s sequential behavior patterns [5], [6], [7]. For example, Factorized Personalized Markov Chain (FPMC) [5] is a representative work which combines user preference and user’s location transition pattern to perform sequential recommendation. To investigate the high-order sequential influence, the recent mainstream techniques often resort to neural network based methods such as recurrent neural network (RNN) [8] and convolutional neural network (CNN) [9]. In general, recurrent models [10], [11] summarize historical interactions between consecutive POIs via a hidden state and update the hidden state during training steps for next POI recommendation. As to convolutional models [3], [4], word embedding techniques like word2vec [6] are widely adopted to obtain a compact representation of sparse sequential check-in data, and then the model apply convolution operation to extract transition patterns by sliding parameterized transformations across the input sequences.
Despite their success, the major drawback of those methods is that they ignore the real-time demand of users, which can be derived from various context information associated with historical POIs. On the one hand, a user’s real-time demand is highly dependent on his current states, such as the current time and his/her geographical position. For example, a user’s demand at midnoon is usually a restaurant. On the other hand, a user’s real-time demand is influenced by his previous visited POIs, including the first-order and high-order previous POI(s). As illustrated in Fig. 1, there is a check-in list {airport, restaurant, hotel}, a user’s demand to visit a hotel is highly influenced by its second-order previous POI (i.e., the airport). Furthermore, the influence of first-order previously visited POI for next visiting POI is apparent. In other words, a user is unlikely to revisit a restaurant right after his last visit. As a result, a user’s decision on where to go next is actually a trade-off between his preference and real-time demand. Exploiting the user preference and real-time demand simultaneously thus offer a more natural way to model the user’s visiting behaviors.
In this paper, we present a new next POI recommendation method, called DSPR, by exploring user’s preferences and real-time demand simultaneously. To model the real-time demand, different kinds of contextual information are exploited. Specifically, user’s real-time demand is modeled from four aspects: including (1) temporal-constraint driven demand such as absolute time and date or time duration; (2) Geographic–constraint driven demand like POI–POI transition distance; (3) POI–POI transition time; (4) Semantic Category of POI, which reveals the functionality of the POI. To model user’s preferences to POIs, weighted matrix factorization is employed. Building upon these, an attention-based recurrent neural network is further introduced to capture the importance of user preferences and different aspects of real-time demand. The final decision vector is derived as a weighted sum of user preferences and different aspects of real-time demand. Specifically, DSPR first represents each POI in a check-in sequence as a POI vector that captures both user preference and real-time demand. Then a decision vector is derived from the attention based recurrent neural network. The decision vector captures the sequential pattern and assigns weights automatically to user’s real-time demands and user-POI preference. In the end, a fully connected neural network takes the POI vector and decision vector as input to perform next POI recommendation, higher consistency between these two vectors produces higher recommendation scores.
By considering both user’s preference and real-time demands, our major contributions in this paper can be summarized as follows.
- •
We propose a new sequential POI recommendation method, which considers users’ general preferences and real-time demands simultaneously for personalized next POI recommendation. To the best of our knowledge, this is the first work that models users’ real-time demand for next POI recommendation.
- •
We represent users’ real-time demand and general preferences as latent vectors, and employ an attention-based recurrent neural network model, which learns the importance of different aspects of demands and general preference automatically, to capture the high-order influence among POIs
- •
We perform experiments on three large-scale LBSN datasets, the results demonstrate that DSPR outperforms many state-of-art next POI recommendation approaches.
The rest of the paper is organized as follows. In Section 2, we review some related works. Section 3 formally defines the problem and describes the DSPR framework. Section 4 shows the model training. We report the experimental results in Section 5. Finally, Section 6 gives our conclusion and future work.
Section snippets
Related work
In the past ten years, numerous next POI recommendation algorithms have been proposed (e.g., [5], [12], [13], to mention a few). Here, we only highlight some most closely related works, including conventional POI Recommendation, side information for POI recommendation, and sequence-aware POI recommendation.
Problem statement
Let denote the set of users, and denote the set of POIs. Each POI can be identified by a unique geographical location vector , where lon denotes the longitude and lat the latitude, respectively. Each POI belongs to a category set , e.g., nightlife spot, restaurant, fitness center, etc. Definition 1 (Check-in) A check-in record is a tuple , where is the absolute time when the check-in happens. Definition 2 (Check-in Sequence
Model training
The objective of DSPR model is to maximize the output probability of the ground truth POI and minimize the probability of negative POIs. To achieve this goal, we adapt negative sampling [45] to randomly select several unseen POIs in to train the model. We use binary cross-entropy as the loss function for DSPR. Given the training sequence set of all users, the objective function is defined as follows.where is the
Experiment
Datasets. To evaluate our proposed method, we perform experiments on three public datasets, including Gowalla2 [47], Foursquare data provided in [48], and Tencent data obtained from [49]. All three datasets are very sparse. For Gowalla and Foursquare, we process the raw data by removing inactive users who have visited less than 15 POIs. Similarly, inactive POIs that are visited less than 15 times are removed from the two data sets. As for Tencent
Conclusion
In this study, we propose a new sequential point-of-interest (POI) recommendation method, called DSPR, by exploiting user’s preferences and real-time demand simultaneously. To model the real-time demand, different contextual information is exploited, including the category of POIs, absolute time, POI–POI transition distance, POI–POI transition-time. Further, POIs are represented as POI vectors by incorporating the user’s preference and these contextual factors. POI vectors are further modeled
CRediT authorship contribution statement
Honglian Wang: Writing - original draft, Conceptualization, Methodology, Investigation, Software. Peiyan Li: Writing - review & editing. Yang Liu: Writing - review & editing. Junming Shao: Writing - review & editing, Conceptualization, Methodology, Supervision.
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 is supported by the National Natural Science Foundation of China (61976044, 41601025), Fundamental Research Funds for the Central Universities (ZYGX2019Z014), Fok Ying-Tong Education Foundation for Young Teachers in the Higher Education Institutions of China (161062), National key research and development program (2016YFB0502300, 2018YFB0804500).
References (49)
- et al.
Multi-objective optimization for location-based and preferences-aware recommendation
Information Sciences
(2020) - et al.
A two-step personalized location recommendation based on multi-objective immune algorithm
Information Sciences
(2019) - et al.
imcrec: A multi-criteria framework for personalized point-of-interest recommendations
Information Sciences
(2019) - et al.
Modeling hierarchical category transition for next poi recommendation with uncertain check-ins
Information Sciences
(2020) - et al.
Geo-teaser: Geo-temporal sequential embedding rank for point-of-interest recommendation
- et al.
Exploiting geographical influence for collaborative point-of-interest recommendation
- et al.
Fused matrix factorization with geographical and social influence in location-based social networks
- et al.
Recnet: a deep neural network for personalized poi recommendation in location-based social networks
International Journal of Geographical Information Science
(2018) - F. Wang, Y. Qu, L. Zheng, C.-T. Lu, S.Y. Philip, Deep and broad learning on content-aware poi recommendation, in:...
- et al.
Factorizing personalized markov chains for next-basket recommendation