Evidential estimation of event locations in microblogs using the Dempster–Shafer theory

https://doi.org/10.1016/j.ipm.2016.06.001Get rights and content

Highlights

  • We estimate locations of events detected in Twitter using Dempster-Shafer theory.

  • Our method combines evidence from multiple tweet features using combination rules.

  • Our method is applicable to any event type and does not require training.

  • Comparisons were made with the Bayesian methods under different settings.

  • Estimations are made for two levels of location granularity with enhanced accuracy.

Abstract

Detecting real-world events by following posts in microblogs has been the motivation of numerous recent studies. In this work, we focus on the spatio-temporal characteristics of events detected in microblogs, and propose a method to estimate their locations using the Dempster–Shafer theory. We utilize three basic location-related features of the posts, namely the latitude-longitude metadata provided by the GPS sensor of the user’s device, the textual content of the post, and the location attribute in the user profile, as three independent sources of evidence. Considering this evidence in a complementary way, we apply combination rules in the Dempster–Shafer theory to fuse them into a single model, and estimate the whereabouts of a detected event. Locations are treated at two levels of granularity, namely, city and town. Using the Dempster–Shafer theory to solve this problem allows uncertainty and missing data to be tolerated, and estimations to be made for sets of locations in terms of upper and lower probabilities. We demonstrate our solution using public tweets on Twitter posted in Turkey. The experimental evaluations conducted on a wide range of events including earthquakes, sports, weather, and street protests indicate higher success rates than the existing state of the art methods.

Introduction

Social networks, particularly microblogs, are increasingly used as communication platforms by millions of people worldwide (Java, Song, Finin, & Tseng, 2007). The availability of Internet access and affordable smart phones allows people to easily access online services to share photos, videos, and text messages with others in their network. With its innovative microblogging concept, Twitter is among the most popular social networking platforms enabling users to post 140 character text messages as tweets, and share them with their followers.

Tweets are usually very responsive to the real-world events, sometimes even more immediate than the relay of traditional news resources (Imran, Castillo, Diaz, & Vieweg, 2015). This has led to Topic Detection and Tracking (TDT) techniques that were previously proposed for newspaper articles and blog posts being extended and adapted to perform event detection on Twitter (Allan, 2002, Atefeh, Khreich, 2015, Fiscus, Doddington, 2002, Yang, Pierce, Carbonell, 1998). In these studies, an event is defined as an activity that happens at a specific time and place. Although state of the art solutions detect events and their times with satisfactory accuracy, we believe that the current solutions about the “place” are still in infancy. In this work, we particularly focus on the location aspects of events in microblogs. We use Twitter as a representative of microblogs, due to its widespread usage and easily accessible public data. However, we believe our proposed solution is also adaptable to other microblogs.

Tweets can contain geographical footprints (Stefanidis, Crooks, & Radzikowski, 2013). A tweet becomes geotagged in terms of latitude and longitude if the user posts it using a GPS-enabled device and allows the sharing of this geographic information. Moreover, users may mention places in the content of their tweets, or state their home location in their Twitter profiles. These three attributes of tweets are widely used for their spatial analysis. However, each attribute poses its own challenge. For example, although GPS coordinates provide a precise geographic position on earth in terms of latitude and longitude, this location and that of the event mentioned in the tweet may not be the same. Once an event is broadcast in the media or shared among people via phone calls or retweets, the GPS coordinates of the recently posted tweets may quickly spread across locations that are further away, making the event location more difficult to detect (Crooks, Croitoru, Stefanidis, Radzikowski, 2013, Sakaki, Okazaki, Matsuo, 2013).

Challenges related to the tweet content and location attribute of user profiles mostly concern text processing and geoparsing, i.e., relating a given text to spatial locations (Hill, 2006). These attributes are uncontrolled free-text fields. Their quality of content is not as good as that contained in news articles due to the idiosyncratic spellings, unusual writing conventions and abbreviations. Therefore, state of the art NLP parsers do not perform as accurately on tweets (Lingad, Karimi, & Yin, 2013).

In addition to these challenges, users do not have to reveal their GPS locations, or their city of residence in their profiles. They also do not have to mention any location name in their tweets. As a result, uncertainty and a lack of rich and reliable data is a major common problem to be overcome.

Problem definition: Spatial analysis in social networks may focus on several aspects. Some studies aim to infer the location where a photo or tweet has been posted even if the user did not share the GPS data of the mobile device (Lee, Ganti, Srivatsa, Liu, 2014, Van Laere, Schockaert, Dhoedt, 2012, Watanabe, Ochi, Okabe, Onai, 2011). Similarly, there are efforts to assign geographical coordinates to textual resources, such as Wikipedia articles (Van Laere, Schockaert, Tanasescu, Dhoedt, & Jones, 2014). Estimating locations of users by utilizing the content in social networks is another active research area (Chang, Lee, Eltaher, Lee, 2012, Cheng, Caverlee, Lee, 2010). In this work, our geospatial analysis focuses on estimating locations of events that can be detected by examining microposts in social networks. Therefore, we describe the problem as follows; given a set of tweets that are presumably about an event, estimate the location of this event by exploiting the available evidence in the tweets and in the profiles of the users who posted the tweets.

This problem is also referred to as event localization by Giridhar, Abdelzaher, George, and Kaplan (2015a); Giridhar et al. (2015b). As the definition of an event, we adopt the terminology in the literature, and define it as an activity that happens at a specific time and place (Allan, 2002, Atefeh, Khreich, 2015, Fiscus, Doddington, 2002). Moreover, in this work, we assume that an event might be happening at multiple locations. For example, a hurricane hitting several cities on the coast can be considered as an event with multiple locations.

Contributions: According to our observation, in most of the similar previous studies, the three location-related attributes of the tweets are considered alternatives to each other. If tweets are expected to have been sent from places closer to the event location, then the GPS coordinates of the geotagged tweets might be a fruitful source of information (Crooks, Croitoru, Stefanidis, Radzikowski, 2013, Sakaki, Okazaki, Matsuo, 2010). For the tweets that are not geotagged, the location attribute in the user profile can give an insight into the location of the event (Achrekar, Gandhe, Lazarus, Yu, Liu, 2013, Sakaki, Okazaki, Matsuo, 2013). Alternatively, if we postulate references to the event location in tweets, then the tweet content can be useful for location estimation (Unankard, Li, & Sharaf, 2015). We believe that the location-related features in tweets are not alternatives for each other, but rather, they complement each other. In this work, we use all three spatial attributes in tweets as evidence sources and combine them systematically in a single model in order to estimate the locations of events, represented by collections of clustered tweets. To this end, we propose applying the theory of belief functions, also called the Dempster–Shafer (DS) theory (Dempster, 1967, Dempster, 1968, Shafer, 1976). The contributions of this work can be summarized as follows:

  • The problem of location estimation for events is investigated using DS theory, which allows us to use the existing evidence pertaining to the location of event in a complementary way and extract belief intervals for the candidate locations.

  • The proposed location estimation method is not specific to the event type thus, it does not require any prior event annotation for training. It is experimentally evaluated on a set of tweets posted in Turkey about events of different types, including concerts, sports, street protests, accidents, and earthquakes. The results show that the proposed method can estimate the location of events with higher accuracy in comparison to the existing state of the art methods.

  • Estimations are made for locations at multiple granularities, namely at the city and town levels. Accordingly, we define an association of evidence between coarse-grained and fine-grained data based on the mixed class hypothesis in DS theory.

  • We demonstrate that the contribution of each attribute in the location estimation problem may change temporally depending on the event type. For some events, GPS coordinates are very reliable for the first few tweets, but they diffuse over time as more tweets are received. For some other types of events, the location references in the tweet content turns out to be more accurate source of evidence over time.

  • Since DS theory yields probability intervals for each discrete geographical entity in the domain, all the locations related with a given event are marked accordingly on a map. This view offers an intuitive graphic representation of the geography of the event.

We presented a preliminary version of the approach in Ozdikis, Oguztuzun, and Karagoz (2013). In the current work, the description of the approach and experimental results are considerably enhanced and extended. We enrich our solution to cover various types of events, and analyze the results of the proposed methods with respect to their varying spatial characteristics. We implement three combination rules in DS theory, which differ in the way of handling conflicts in tweet features. Moreover, we estimate locations at multiple granularities, namely cities and towns, considering a town as part of a city from the viewpoint of settlement hierarchy, and propose a way to define an association of evidence between the coarse-grained and fine-grained locations.

The remainder of this paper is organized as follows. In Sections 2 and 3 we present a review of the literature about the geospatial analysis in social networks, and applications of DS theory, respectively. Section 4 is devoted to our evidential location estimation method using tweets related to events. The results of the evaluations and comparison with the baseline methods are discussed in Section 5. Lastly, we present our concluding remarks in Section 6.

Section snippets

Geospatial analysis on social networks

Recently, numerous studies have been conducted to detect real-world events by collecting public tweets in Twitter (Atefeh & Khreich, 2015). These include methods to identify earthquakes (Crooks, Croitoru, Stefanidis, Radzikowski, 2013, Sakaki, Okazaki, Matsuo, 2010, Sakaki, Okazaki, Matsuo, 2013), disasters and crises (Yin, Lampert, Cameron, Robinson, & Power, 2012), sports events (Corney, Martin, & Göker, 2014), epidemics and diseases (Achrekar, Gandhe, Lazarus, Yu, Liu, 2013, Padmanabhan,

Dempster–Shafer theory

Arthur P. Dempster (Dempster, 1967, Dempster, 1968) introduced a generalized Bayesian inference model based on evidence and evidential reasoning which was further extended by the work of Glenn Shafer (Shafer, 1976) and is known as the Dempster–Shafer (DS) theory. The first step in applying this theory to a problem is defining the propositional space of possible solutions, the Frame of Discernment, denoted by Θ. In the case of location estimation, this set is composed of all the locations (e.g.,

Evidential location estimation

In this section we introduce the way in which belief-plausibility intervals are assigned to the locations for a given set of tweets about an event, and finally make a location estimation for the event using these assignments. We describe locations as cities and their towns in a country, organized in a geographical settlement hierarchy in which the country is divided into cities which are further divided into towns. The information about the cities and towns is stored as a gazetteer that

Evaluation and discussion

In this section, we present the results of our evidential location estimation algorithm applied to a group of tweet sets corresponding to various events, and compare them with the results of baseline methods. We confine our problem domain to include the tweets, events, and locations in Turkey. Therefore, we aggregated the definitions of 81 cities and 964 towns in Turkey from open data sources on the web, namely the OpenStreetMap2 and Wikimapia,3

Conclusion

In this work, we propose an unsupervised method based on the Dempster–Shafer (DS) theory for the location estimation problem for events detected in microblogs, and apply it to Twitter. We use three location-related features in the tweets as evidence sources, represent the evidence by three separate mass functions, and combine them into a single solution using combination rules in DS theory. The advantages of fusing these features in a single model are twofold; first, we can leverage all of

Acknowledgment

This work is supported by TUBITAK under grant number 112E275 and ICT COST Action IC1203. The authors are grateful to Mr. Özcan Dülger for his support in the filter implementations.

References (61)

  • A. Al-Ani et al.

    A new technique for combining multiple classifiers using the Dempster-Shafer theory of evidence

    Journal of Artificial Intelligence Research

    (2002)
  • E. Amitay et al.

    Web-a-where: Geotagging web content

    Proceedings of ACM SIGIR 2004

    (2004)
  • F. Atefeh et al.

    A survey of techniques for event detection in Twitter

    Computational Intelligence

    (2015)
  • S.S. Blackman et al.

    Design and Analysis of Modern Tracking Systems

    (1999)
  • ChangH.-w. et al.

    @Phillies tweeting from Philly? predicting Twitter user locations with spatial word usage

    Proceedings of ASONAM 2012

    (2012)
  • ChengZ. et al.

    You are where you tweet: A content-based approach to geo-locating Twitter users

    Proceedings of CIKM 2010

    (2010)
  • K. Coombs et al.

    Using Dempster-Shafer methods for object classification in the theater ballistic missile environment

    Proceedings of SPIE, sensor fusion: Architectures, algorithms, and applications III

    (1999)
  • D. Corney et al.

    Spot the ball: Detecting sports events on Twitter

  • A. Crooks et al.

    #Earthquake: Twitter as a distributed sensor system

    Transactions in GIS

    (2013)
  • S.P. Curley

    The application of Dempster-Shafer theory demonstrated with justification provided by legal evidence

    Judgment and Decision Making

    (2007)
  • A.P. Dempster

    Upper and lower probabilities induced by a multivalued mapping

    The Annals of Mathematical Statistics

    (1967)
  • A.P. Dempster

    A generalization of Bayesian inference

    Journal of the Royal Statistical Society. Series B (Methodological)

    (1968)
  • D. Dubois et al.

    Representation and combination of uncertainty with belief functions and possibility measures

    Computational Intelligence

    (1988)
  • J.G. Fiscus et al.

    Topic detection and tracking evaluation overview

  • D. Fox et al.

    Bayesian filtering for location estimation

    Pervasive Computing, IEEE

    (2003)
  • FuC. et al.

    The group consensus based evidential reasoning approach for multiple attributive group decision analysis

    European Journal of Operational Research

    (2010)
  • P. Giridhar et al.

    Joint localization of events and sources in social networks

    Distributed computing in sensor systems (DCOSS)

    (2015)
  • D. He et al.

    Combining evidence for automatic web session identification

    Information Processing and Management

    (2002)
  • L.L. Hill

    Georeferencing: The Geographic Associations of Information

    (2006)
  • Cited by (29)

    • Extracting temporal and causal relations based on event networks

      2020, Information Processing and Management
    • A Google Trends spatial clustering approach for a worldwide Twitter user geolocation

      2020, Information Processing and Management
      Citation Excerpt :

      In contrast with several state-of-the-art works (e.g., Celik & Dokuz, 2018; Do, Nguyen, Tsiligianni, Cornelis, & Deligiannis, 2018; Huang & Carley, 2019; Ozdikis et al., 2019; Paule et al., 2019), a pure unsupervised WD approach is adopted and thus no geographic labeled data (e.g., tweets or user location profiles) is required, only historical tweet nouns and GT data. Moreover, we do not use LIW, as adopted in (Bakerman et al., 2018; Ozdikis et al., 2016; Shahraki, Fatemi, & Malazi, 2019; Williams et al., 2017), since LIW often assumes finite and rather static set of locations, typically associated to small world regions. Instead, we use tweet nouns, which can be dynamically updated and that can refer to geographic words and also other terms with a location context (e.g., events, people or organizations).

    • Data-Driven Computational Social Science: A Survey

      2020, Big Data Research
      Citation Excerpt :

      how does human relationship evolve? how to detect human community from social media data?) [9,10]. Scholars in this emerging area can use related data to solve these social problems with advanced computing technologies.

    View all citing articles on Scopus
    View full text