Real-time event embedding for POI recommendation

Abstract

Location-based social networks (LBSNs) allow users to check-in and share daily lives with others. We have witnessed very rapid development of LBSNs in recent years. Point-of-Interest (POI) recommendation is one of the core services in LBSNs. In this study, we propose a real-time POI embedding model. Instead of capturing intrinsic information, the proposed approach is able to mine real-time information of places and learn the latent representations according to the corresponding geo-tagged posts. On one hand, we employ a Convolutional Neural Networks (CNN) to mine textual information of POIs and learn their intrinsic representation. On the other hand, a multimodal embedding model of location, time and text is applied to keep monitoring posts on POIs and extracts a set of features for representing events or burst information that may attract users. Furthermore, we combine real-time POI embedding with matrix factorization method and propose a more comprehensive POI recommendation algorithm. To verify the effectiveness of our proposed method, we conduct experiments on Twitter dataset with geo-tagged tweets in NYC. Experimental results show that POI recommendation system with taking real-time event into consideration can strongly improve the performance than the one without.

Introduction

In recent years, social networking is widespread growing throughout the world. According to the statistics revealed on Statista,¹ more and more people tend to share their daily life with friends and family, which has become a city phenomenon. Several famous location-based social networks (LBSNs), such as Facebook and Twitter, provide geo-tagging (check-in) services, in which users can share their activities and locations (Fig. 1). For the fast grown-up of LBSNs, point-of-interest (POI) recommendation has become a hot issue to help users discover interesting locations by mining users’ preference from a large number of check-in behaviors. However, POI recommendation is challenging as the number of POIs is often so large compared to the number of POIs visited by an individual, resulting in a very sparse user-POI matrix. Many researchers focus on studying sequential check-in to learn users’ preference [30], [4], [31], while ignoring the POI information. Recent works show integrating textual information on POI can improve the performance of POI recommendation. For instance, Liu et al. [13] recommended POIs by mining their textual information to topics with Latent Dirichlet allocation (LDA). [26] mined latent features of POIs with CNN. While those works mostly focus on embedding intrinsic information of POI, two major factors are being ignored. One is the factor of time. In a scenario that given the user and the current time, we want to recommend a POI which is most suitable for him/her at that time. The other one is the factor of event. We assume that event will directly affect users’ desire to visit one place even though they are not supposed to visit based on their preference.

For example, a Coldplay fan who seldom has sports checks-in at a stadium because of the Coldplay concert. However, most works on POI recommender that exploit temporal characteristics do not consider the factor of event. Therefore, we are going to develop a real-time POI embedding metric that consider both intrinsic and time-aware information of POI and employ it to improve POI recommendation.

We assume that instead of eye-catching POI, users would prefer a place with an event which they are interested in. To observe the importance of discovering events, an empirical data analysis is applied to answer the following two questions: (1) how is the population of POIs behaving whether events happened or not? (2) How the posts on POIs perform with events happened? We analyze a specific POI with its geo-tagged posts in sequential days, as shown in Fig. 2. We divide days into hours and observed the (check-ins) records as well as the contents of posts. We also analyze the high occurrence of keywords from tweets which were posted on an event day, as shown in Fig. 3. We sum up with two obvious phenomena, (1) the number of check-ins apparently increase when events happened; meanwhile (2) from the content of related posts, we find that abnormal vocabs would be mentioned more frequently, hence the words with high frequency could well describe the event.

It is clear that geo-tagged posts provide immediate and direct information of events. For the above-mentioned reasons, this study proposes a real-time POI recommendation system with event detection to analyze textual information on social media. Specifically, our POI recommender consists of two major characteristics:

• Time-aware POI Information which capturing both intrinsic and time-aware information on POIs.

• Real-time Event Tracking on social media for extracting latest and highly attentive information on POIs.

For the intrinsic representation of POIs, we develop a CNN metric in mining textual description or reviews to capture intrinsic features for POIs, inspiring by [26]. Whereas for the real-time process, we construct a multimodal embedding of keyword, location and time to explore instant semantic representation of geo-tagged posts given the POI, so as to keep track of whether there are events occurring. Our recommendation system combines real-time embedding to matrix factorization.

The thesis is organized as follows: Section 2 reviews the related works; Section 3 describes the details of the proposed model. Section 4 describes the experiment to evaluate the performance of our model, and finally, Section 5 closes the thesis with conclusions and suggestions for future work.