Small-scale events are emerging as attractive objects of research. On Twitter, small-scale events represent weak sensors that report things happening in specific times and places. While previous work addressed the issue of detecting such events, very little is known so far about their inherent properties. In this paper, our main objective was to analyse the spatio-temporal peculiarities of small-scale events w.r.t different levels of location granularity, and to understand the general trend of their propagation along their lifetimes. Our findings suggest that (1) users involved in small-scale events mostly gravitate not significantly far from the geographical focus; (2) events do not exhibit major peaks; and (3) there exists distinct events that we can identify from users’ posts that significantly differ from topic distribution, focus concentration and propagation distance perspectives across time.

CCS CONCEPTS

• Computer systems organization →Embedded systems; Redundancy; Robotics; • Networks →Network reliability;

KEYWORDS

Small-scale event; geo-tagged tweets; focus; entropy

ACM Reference Format: Paul Mousset, Yoann Pitarch, and Lynda Tamine. 2018. Studying the Spatio- Temporal Dynamics of Small-Scale Events in Twitter. In HT ’18: 29th ACM

Conference on Hypertext and Social Media, July 9–12, 2018, Baltimore, MD, USA. ACM, New York, NY, USA, 9 pages. https://doi.org/10.1145/3209542. 3209561

Microblogging platforms such as Twitter provide active communication channels and a gold-mine of timely real-world information which has been shown to be highly effective for gaining knowledge about people’s profiles [22], and opinions [16] to cite just a few. In particular, for events such as festivals, political campaigns, pandemics and crisis situations, user-generated micro-posts play a crucial role as social sensors by allowing the monitoring of users’ activities and the provision of timely responses and

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. HT ’18, July 9–12, 2018, Baltimore, MD, USA © 2018 Association for Computing Machinery. ACM ISBN 978-1-4503-5427-1/18/07...$15.00 https://doi.org/10.1145/3209542.3209561

recommendations [12]. More specifically, event-related tweet streams provide valuable spatio-temporal data such as text messages with location mentions, the timestamp of the post and the geolocation of the user who posted the tweet. With the increasing connectivity of users through wireless networks and the wide use of mobile devices, geo-tagged tweets are currently created daily. This phenomena allowed the intensive use of Twitter data for detecting and monitoring both large-scale events (eg., earthquakes and epidemics) and small-scale events (eg., festivals, crimes and protests). An important body of early research focused on detecting, analysing and mining behavioural patterns from large-scale events (also called global events), which are bursty in the entire stream, impact a wide spatial area and trigger an important audience [6, 23]. Recently, there has been a growing research interest in detecting [1, 26, 30–32] and analysing [24, 25, 30] small-scale events. Unlike large-scale events, small-scale events (also called local or localized events) are generally micro-phenomena that stimulate people to post a low number of messages for a certain period of time in a local region. Such events play the roles of weak signals which have potential in several applications such as public order protection and traffic road assistance. However, the literature review reveals that very little has been understood so far about the spatio-temporal dynamics of such events [24, 25, 30]. A prior study [30] focused on the analysis of user physical network structure during two micro-events, namely a parking garage collapse in Atlanta and a church shooting in Wichita. The results mainly revealed that the event-related structure of the networks is not particularly more dense than the Twitter network structure and that central Twitterers are geographically central particularly in more spatially narrowed events. In [24, 25], authors examined the users’ posts during two incidents which refer to small-scale events that result in damage or injuries. The authors found that different types of users (eg., journalists, organisation and citizens) report on the incidents and that citizens are generally faster than official sources in propagating tweet posts.

In this paper we pursue this line of research and report the findings of analyses that are designed to investigate the spatio-temporal dynamics of small-scale events. By using the focus and entropy measures, we thoroughly study the spatial and timely tweet post distributions of such events based on a wide set of event types automatically identified in geo-tagged tweet streams. Moreover, we consider locations at varying levels of granularity, from the borough to the Point Of Interest (POI) level. The key differences between close previous work [24, 25, 30] and ours are the following: (1) previous work focused on the analysis and comparison of network structure in the Twitter network vs. event-related network [30] and

the identification of user types involved in the event [24, 25]. Moreover, the studies used limited samples of prior specific events (eg., incidents) in terms of number of tweets (655 event-related tweets used in [24, 25]), as well as number of events (2 events studied in [30]). In contrast, we deeply analyse the spatial narrowness of users involved in the events and study the propagation trend of their posts w.r.t varying levels of location; furthermore, our study relies on a significant number of event-related tweets (22832) as well as a significant number of events (410) belonging to a set of event types automatically identified from the users’ posts. (2) We also examine the spatio-temporal evolution of small-scale events along their lifetime while the temporal dimension has not been addressed in previous work [24, 25, 30]. More specifically, our study is mainly designed for answering the following general questions: What is a rough estimate of the geographical impact of small-scale events in terms of situationalawarness among users? How do small-scale events evolve in time and space? Answering to these questions has many potential applications in improving the designing of better web tracking and searching services and helping both organisation and citizen better fulfill their information needs. For example, modelling the timely propagation of small-scale events from their center of gravity enables the development of more robust user’s location estimators over time which can be useful for fine-grained location-based personalized information services [10] and designing more effective monitoring services allowing to better plan social events and offer user facilities [3]. A better understanding of their evolving spatio-temporal dispersion enables richer retrieval models answering location-based queries (eg., What is happening at specific location X?) that could improve the spatial extent of users’ awarness during crisis situations. The analyses performed in our work directly address the following research questions:

RQ1: What is the level of users’ narrowness while posting small-scale event-related tweets? RQ2: How do location and time bound tweet propagation during small-scale events? RQ3: Can we identify and characterize event types with the tweet publications collected from users? The remainder of this paper is organized as follows: Section 2 presents the related work. We detail the data and methods that are used in our analyses in Section 3. In Section 4, we report the findings of our study w.r.t the aforementioned research questions. We discuss the implications of our findings and conclude in Section 5.

2 RELATED WORK 2.1 Large-scale vs. small-scale event detection

To the best of our knowledge, there is no consensual definition of an ’event’ [20, 24]. One widely used definition has been introduced by the Topic Detection and Tracking (TDT) project which defines an event as ’an unique thing that happens at some point in time’ [13]. The main common facets of an event are time, place and audience as introduced by authors in [7] who define an event as a real-world occurrence with an associated time period, time-ordered messages discussing the occurrence during a period of time. According to spatial and social impacts, we can distinguish between large-scale events such as earthquakes and epidemics (also called global events)

and small-scale events such as crims, protests and festivals (also called incidents, local or localized events). While the former give rise to massive user-generated content and impact a wide spatial region, the latter lead to the posting of a low amount of content within a small region. A number of early research work have investigated global event detection in Twitter [6, 23]. Related approaches fall in the documentbased and feature-based categories. The key idea of document-based approaches relies on the association between document cluster based on a shared topic and the notion of event [2]. For instance, authors in [2] built events as clusters based on content similarity and user proximity. In feature-based approaches [11], the event detection algorithm rather relies on bursty features such as keywords and phrases. Such burstiness is captured as event signal wich is filtered and transformed (eg., using time-series) to identify events. The high connectivity of users through wireless networks and the wide availability of geo-tagged tweets nowadays gave rise to a recent emerging interest toward small-scale event detection which is particularly challenging given the low amount of induced posts [1, 26, 32]. The driving idea behind small-scale event detection algorithms is to identify joint geo-spatial and topical correlations in Twitter streams. The main underlying assumption is that, topical cohesion of even limited samples of tweets posted in a narrow geographical space, is a weak signal of event occurrence. Based on this general assumption, authors in [1] proposed the EvenTweet algorithm which detects small-scale events by running four main stages: (1) determining time windows of bursty words; (2) identifying localized words by computing spatial entropy; (3) spatial clustering of localized words; and (4) associating each cluster with an event and then ranking events according to spatial coverage and burstiness measures. In the same spirit, the GeoBurst algorithm [32] relies on three main steps to detect small-scale events: (1) identifies representative tweets (called pivots) and associated geo-topic correlated tweets in slicing time windows; (2) compares the clusters of tweets to historical streaming tweets to identify bursty clusters of tweets among the candidate ones identified in the first stage; and (3) further identifies new pivots in successive windows.

2.2 Analysing spatio-temporal Twitter data

The prevalence of social media services such as Facebook and Twitter and the increasing use of mobile devices, have enabled the availability of a powerful source of streamed user-generated content about location and regional human behaviour. This motivates the spatio-temporal analysis of social media data for various purposes such as crisis management [21] and future prediction [5]. On Twitter, the literature review reveals that a large body of studies investigated the analysis of spatio-temporal metadata for different purposes [4, 14, 18, 27]. A first category of work have analysed the spatio-temporal properties of tweets to better understand their diffusion [4, 14] or discover spatio-temporal dependent topics that are addressed by Twitter users. For example, in [14], the authors mainly observed the joint phenomena of the high-locality of hashtags and their high-speed propagation over time. Another category of work have used spatio-temporal data from Twitter to design web services such as automatic summarizers [19] or monitoring systems [3]. For instance, in [19], the authors developed an event summarizer called CEST that exploits Twitter

geographical metadata mainly including POI. Once the event is detected, the CEST system provides a high-level picture of the spatial extent of the event as well as the distribution of positive vs. negative opinions embedded in the event-related tweets. An other important category of work focused on mining from largescale or small-scale events. Early work studied large-scale events through the examination of the regional trends and temporal patterns of users’ behaviour to address medical concerns [18] or better manage emergency situations [27]. Lee et al. [18] designed a surveillance system for early prediction of seasonal disease outbreaks such as flu. Using tweet analytics about the timeline and geographical distribution of disease symptoms, the system can facilitate the monitoring of health resource allocation during epidemics. In [27], the authors analysed users’ micro-posts across two disaster events that took place in the US: the Red River Floods and the Oklahoma Grassfires. They outlined that unlike in the overall tweet stream, there were a high proportion (more than 78%) of both geo-tagged tweets and location mentions in tweet texts during natural disasters which suggests that users are aware of the importance of geo-location information in collaboratively managing an emergency. Moreover, the analysis of the event time-line showed a clear picture of the spread of situational awareness among Twitterers. Other work, more close to ours focused on small-scale events, also called incidents in the specific case where they induce damage and injuries [24, 25, 30]. Their analysis mainly concerned the study of the impact of such events on the network structure [30] and the categorisation of user profiles (eg., organization, citizen) according to predefined event types such as crash and fire [24, 25]. The findings revealed that influencial users kept their roles in event-related networks [30] and that citizen were more importantly involved than officials in the event spread [24, 25].

3 STUDY DESIGN 3.1 Data

3.1.1 Twitter datasets. We analysed two collections of geotagged tweets related to New York City, and restricted our study to English-language tweets. The geotag provides latitude-longitude coordinates about the physical location of the associated Twitter user. To facilitate the analysis of multiple levels of location granularity, we used the latitude-longitude coordinates and mapped them to the borough and neighbourhood and to the most likely Place Of Interest (POI). To perform this mapping, we used a recent state-of-the art POI annotation method that relies on geo-tagged tweets [33]. We have particularly choosen New York City as the main location for a couple of reasons: (1) given the high number and the diversity of event types that might occur in New York; (2) the availability of rich resources that provide POI descriptions in New York City. The characteristics of the datasets are the following.

• The first dataset, which is referred to as NY2014, was released by Zhang et al. [32]. It consists of a sample of 2.4 million of tweets that were geo-tagged in New York and retrieved using the Twitter Streaming API1 from 2014.08.01 to 2014.11.30. • The second dataset was obtained by constantly monitoring the Twitter streams from 2016.06.18 to 2016.12.08 for 116

1https://developer.twitter.com/

cities around the world, which included more than 65 million geo-tagged tweets. For this study, we extracted a sub-dataset, which is referred to as NY2016, that contains only geo-tagged tweets published in New York from 2016.10.01 to 2016.11.30. It contains approximately 2.98 million tweets.

3.1.2 Event labelling. The process of building a pool of smallscale events from the two datasets consisted in a 2-step task:

• Step 1: Generating candidate small-scale events. The objective of this step was to extract from the Twitter datasets clusters of representative event-related tweets. To achieve this objective without bias induced by the event detection algorithm, we used two state-of-the art event detection algorithms, namely the EvenTweet [1] and GeoBurst algorithms [32]. In addition to our confidence in their effectiveness and efficiency, the main advantage of these algorithms is that they provide clusters of tweets and/or authoritative words that facilitate further human annotation. The generation of candidate small-scale events was run as follows: (1) Apply separately each of the EvenTweet and GeoBurst

algorithm on each dataset, namely the NY2014 and NY2016 datasets. More specifically: (i) we ran the GeoBurst algorithm using very similar parameters to those used by Zhang et al. [32]. More precisely, the kernel bandwith h = 0.01, the ranking parameter for balancing spatial and temporal burstiness η = 0.5 and the RWR similarity threshold δ = 0.01 with sliding time windows of 6 hours, including 3 hours of overlapping time. (ii) we ran the EvenTweet algorithm based on the partitioning of the whole space into NxN small grids with N = 50. At the end of this stage we obtained a pair of ranked lists of word clusters for each dataset. (2) With respect to our goal which consists in avoiding

the bias induced by the event detection algorithm on the spatio-temporal properties of the tweets, we only selected a subset of common events that were detected by both algorithms as relevant candidate events. Instead of applying exact matching which is highly unlikely given the difference in the techniques used by the GeoBurst and EvenTweet algorithms for tweet clustering, we applied an approximate matching based on a clustering similarity constrained by a threshold: the Szymkiewicz-Simpson overlap coefficient with a threshold set up to 0.5. This first step resulted in 1163 and 1802 event clusters for the NY2014 and NY2016 datasets respectively.

• Step 2: Building the ground truth event datasets. In the previous step we obtained a set of candidate small-scale events. However, given that the precision of each of the GeoBurst and EvenTweet is less than 100%, human annotation was required to build the gold dataset. To this end, we set up a crowdsourced annotation tasks using CrowdFlower2, which is a popular crowdsourcing platform. The objective of this task was to build the gold set of small-scale events from the candidate events which were represented as clusters of

2http://www.crowdflower.com/

Table 1: Descriptive statistics of the NY2014 and NY2016 datasets.

NY2014 NY2016

In this paper, we analysed the spatio-temporal dynamics of smallscale events. Our primary objective was to determine the perimeters of their geographical social impacts at different levels of location granularity, and to gain understanding of their audience and the general trends of their propagation along their lifetimes. Our results suggest the following trends:

• In response to RQ1, the results show that the focus is a significant origin location from which users post their tweets, particularly at coarser levels of location granularity. Moreover, even if events propagate over several locations, they mostly reach narrow regions. Building on these findings, one relevant practical implication that we envision is the design of information seeking algorithms that are able to timely and automatically enlarge the event propagation diameter

6Ratio between total number of users and total number of events in the group.

by rooting event mentions to users who are located in narrowed regions. User’s location, if not explicitly provided, could either be inferred using improved state-of-the art algorithms for tweet geo-location [15]. This would increase the situational awareness particularly during security incidents.

• In response to RQ2, we found that the temporal series of events are mostly stable which suggests the absence of significant peaks. We also found that evently timely evolve from diverse locations and quickly stabilize not significantly far from the focus. A relevant research opportunity that arises from this study is to examine these findings alongside previous research findings about large-scale event detection [6] to design novel algorithms that can jointly detect both weak and strong signals in Twitter streams considering appropriate spatio-temporal distribution and density of posts. Such general detectors can provide means for monitoring people’ activities (eg., for public order maintenance purpose).

• In response to RQ3, we found that we can detect distinct types of events with evolutions that are significantly different according to audience, focus concentration and propagation distance trends over time. Based on these findings, the implications for further theoretical investigation is to develop models for predicting event type based on the event-related features. Event type prediction would be a prior step to the development of an automatic visual summarization method that would give a high-level picture of what is happening in a region. Our study has some limitations. First, we only used the focus, entropy and distance metrics to report the analysis results. Although these measures are the primary metrics that are used for the spatiotemporal analysis of events, they are still insufficient for revealing other relevant facets such as propagation rate. Second, enlarging the spatio-temporal scope of our study to other cities and during different periods might give better insights about the generalisability of our findings. This investigation is planned for future work.

This research was supported by IRIT and ATOS research program under ANRT CIFRE grant agreement.

[1] Hamed Abdelhaq, Christian Sengstock, and Michael Gertz. 2013. Eventweet:

Online localized event detection from twitter. PVLDB Endowment 6, 12 (2013), 1326–1329. [2] Charuc C. Aggarwal and Karthik Subbian. 2012. Event Detection in Social Streams.

In Proceedings of the 2012 SIAM International Conference on Data Mining (SDM’12). SIAM, 624–635. [3] Paolo Arcaini, Gloria Bordogna, Dino Ienco, and Simone Sterlacchini. 2016. User-

driven geo-temporal density-based exploration of periodic and not periodic events reported in social networks. Information Sciences 340-341 (2016), 122 – 143. [4] Sebastien Ardon, Amitabha Bagchi, Anirban Mahanti, Amit Ruhela, Aaditeshwar

Seth, Rudra Mohan Tripathy, and Sipat Triukose. 2013. Spatio-temporal and Events Based Analysis of Topic Popularity in Twitter. In The 22nd ACM International Conference on Information and Knowledge Management (CIKM ’13). ACM, 219–228. [5] Sitaram Asur and Bernardo A. Huberman. 2010. Predicting the Future with Social

Media. In Proceedings of the 2010 IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT ’10). IEEE, 492–499. [6] Farzindar Atefeh and Wael Khreich. 2015. A survey of techniques for event

detection in twitter. Computational Intelligence 31, 1 (2015), 132–164. [7] Hila Becker, Mor Naaman, and Luis Gravano. 2011. Beyond Trending Topics: Real-

World Event Identification on Twitter.. In Proceedings of the Fifth International

Conference on Weblogs and Social Media (ICWSM’11). AAAI, 438–441. [8] David M. Blei, Andrew Y. Ng, and Michael I. Jordan. JMLR’03. Latent Dirichlet

Allocation. Journal of Machine Learning Research 3 (JMLR’03), 993–1022. [9] Anders Brodersen, Salvatore Scellato, and Mirjam Wattenhofer. 2012. YouTube

Around the World: Geographic Popularity of Videos. In Proceedings of the 21st World Wide Web Conference 2012 (WWW’12). ACM, 241–250. [10] Zhiyuan Cheng, James Caverlee, and Kyumin Lee. 2010. You Are Where You

Tweet: A Content-based Approach to Geo-locating Twitter Users. In Proceedings of the 19th ACM Conference on Information and Knowledge Management (CIKM ’10). ACM, 759–768. [11] Gabriel Pui Cheong Fung, Jeffrey Xu Yu, Philip S. Yu, and Hongjun Lu. 2005.

Parameter Free Bursty Events Detection in Text Streams. In Proceedings of the 31st International Conference on Very Large Data Bases (VLDB ’05). ACM, 181–192. [12] Bo Hu and Martin Ester. 2013. Spatial Topic Modeling in Online Social Media for

Location Recommendation. In Seventh ACM Conference on Recommender Systems (RecSys ’13). ACM, 25–32. [13] Allan James, Carbonell Jaime, Doddington George, Yamron Jonathan, and Yang

Yiming. 1998. Topic Detection and Tracking Pilot Study Final Report. In DARPA Broadcast News Transcription and Under standing Workshop (BNTUW ’98). Morgan Kaufmann Publishers, 194–218. [14] Krishna Y. Kamath, James Caverlee, Kyumin Lee, and Zhiyuan Cheng. 2013.

Spatio-temporal Dynamics of Online Memes: A Study of Geo-tagged Tweets. In The 22nd International World Wide Web Conference (WWW ’13). ACM, 667–678. [15] Sheila Kinsella, Vanessa Murdock, and Neil O’Hare. 2011. "I’M Eating a Sandwich

in Glasgow": Modeling Locations with Tweets. In Proceedings of the 3rd International CIKM Workshop on Search and Mining User-Generated Contents (SMUC ’11). ACM, 61–68. [16] Efthymios Kouloumpis, Theresa Wilson, and Johanna D Moore. 2011. Twitter

sentiment analysis: The good the bad and the omg! Icwsm 11 (2011), 538–541. [17] Denis Kwiatkowski, Peter C.B. Phillips, Peter Schmidt, and Yongcheol Shin. 1992.

Testing the null hypothesis of stationarity against the alternative of a unit root: How sure are we that economic time series have a unit root? Journal of Econometrics 54, 1 (1992), 159 – 178. [18] Kathy Lee, Ankit Agrawal, and Alok Choudhary. 2013. Real-time Disease Surveil-

lance Using Twitter Data: Demonstration on Flu and Cancer. In The 19th ACM

SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’13). ACM, 1474–1477. [19] Deepa Mallela, Dirk Ahlers, and Maria Soledad Pera. 2017. Mining Twitter

Features for Event Summarization and Rating. In Proceedings of the International Conference on Web Intelligence (WI ’17). ACM, 615–622. [20] Andrew J. McMinn, Yashar Moshfeghi, and Joemon M. Jose. 2013. Building a

Large-scale Corpus for Evaluating Event Detection on Twitter. In The 22nd ACM International Conference on Information and Knowledge Management (CIKM ’13). ACM, 409–418. [21] Stuart E Middleton, Lee Middleton, and Stefano Modafferi. 2014. Real-time crisis

mapping of natural disasters using social media. IEEE Intelligent Systems 29, 2 (2014), 9–17. [22] Alan Mislove, Bimal Viswanath, Krishna P. Gummadi, and Peter Druschel. 2010.

You Are Who You Know: Inferring User Profiles in Online Social Networks. In Proceedings of the Third International Conference on Web Search and Web Data Mining (WSDM ’10). ACM, 251–260. [23] Takeshi Sakaki, Makoto Okazaki, and Yutaka Matsuo. 2010. Earthquake Shakes

Twitter Users: Real-time Event Detection by Social Sensors. In Proceedings of the 19th International Conference on World Wide Web (WWW ’10). ACM, 851–860. [24] Axel Schulz, Eneldo Loza Mencía, and Benedikt Schmidt. 2016. A Rapidprototyping Framework for Extracting Small-scale Incident-related Information in Microblogs. Information Systems 57, C (2016), 88–110. [25] Axel Schulz and Petar Ristoski. 2013. The Car that Hit The Burning House: Under-

standing Small Scale Incident Related Information in Microblogs. In Proceedings of the Seventh International Conference on Weblogs and Social Media (ICWSM’13). AAAI, 11–14. [26] Axel Schulz, Benedikt Schmidt, and Thorsten Strufe. 2015. Small-Scale Incident

Detection Based on Microposts. In Proceedings of the 26th ACM Conference on Hypertext & Social Media (HT’15). ACM, 3–12. [27] Sarah Vieweg, Amanda L. Hughes, Kate Starbird, and Leysia Palen. 2010. Mi-

croblogging During Two Natural Hazards Events: What Twitter May Contribute to Situational Awareness. In Proceedings of the 28th International Conference on Human Factors in Computing Systems (CHI ’10). ACM, 1079–1088. [28] B. L. Welch. 1947. The Generalization of Student’s Problem When Several Differ-

ent Population Variance Are Involved. Biometrika 34, 1-2 (1947), 28–35. [29] David Wilkinson and Mike Thelwall. JASIST’12. Trending Twitter topics in Eng-

lish: An international comparison. Journal of the American Society for Information Science and Technology (JASIST’12), 1631–1646. [30] Sarita Yardi and danah boyd. 2010. Tweeting from the Town Square: Measuring

Geographic Local Networks. In Proceedings of the Fourth International Conference on Weblogs and Social Media (ICWSM’10). AAAI, 194–201. [31] Chao Zhang, Liyuan Liu, Dongming Lei, Quan Yuan, Honglei Zhuang, Tim

Hanratty, and Jiawei Han. 2017. TrioVecEvent: Embedding-Based Online Local Event Detection in Geo-Tagged Tweet Streams. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’17). 595–604. [32] Chao Zhang, Guangyu Zhou, Quan Yuan, Honglei Zhuang, Yu Zheng, Lance

Kaplan, Shaowen Wang, and Jiawei Han. 2016. GeoBurst: Real-Time Local Event Detection in Geo-Tagged Tweet Streams. In Proceedings of the 39th International ACM SIGIR conference on Research and Development in Information Retrieval (SIGIR ’16). ACM, 513–522. [33] Kaiqi Zhao, Gao Cong, and Aixin Sun. 2016. Annotating Points of Interest with

Geo-tagged Tweets. In Proceedings of the 2016 ACM on Conference on Information and Knowledge Management (CIKM’16). ACM, 417–426. [34] Wayne Xin Zhao, Jing Jiang, Jianshu Weng, Jing He, Ee-Peng Lim, Hongfei Yan,

and Xiaoming Li. 2011. Comparing Twitter and Traditional Media Using Topic Models. In Proceedings of the 33rd European Conference on Advances in Information Retrieval (ECIR’11). ACM, 338–349.

The following native-image page facsimiles preserve the source rendering for each identified figure, table, and equation.

Figure 1: Cumulative Distribution Function (CDF) of event focus, entropy and average distance.

Figure 2: Correlation between Focus, Entropy and Distance at POI level.

Figure 3: Ratio of peak occurrence in non-stationary events.

Figure 4: Average event entropy He(t) and focus Fe(t) evolu-

Figure 5: Evolution of the averaged distance and entropy val-

Figure 6: Distribution of number of users and number of tweets w.r.t event groups and topics.

Table 1: Descriptive statistics of the NY2014 and NY2016

Table 2: A sample of small-scale events identified in the

Table 3: Ratio of events w.r.t their duration.

Table 4: Comparison of events with peak vs. without peak.

Equation source facsimile, page 3.

Equation source facsimile, page 4.

Equation source facsimile, page 5.

Equation source facsimile, page 6.

Equation source facsimile, page 7.

The following native IPFS images retain the complete original pages for visual verification.

Source PDF page 1.

Source PDF page 2.

Source PDF page 3.

Source PDF page 4.

Source PDF page 5.

Source PDF page 6.

Source PDF page 7.

Source PDF page 8.

Source PDF page 9.