The recent rise in the use of social networks has resulted in an abundance of information on di erent aspects of everyday social activities that is available online. In the process of analysis of the information originating from social networks, and especially Twitter, an important aspect is that of the geographic coordinates, i.e., geolocalisation, of the relevant information. This information is used by a variety of applications for the better understanding of an urban area, the tracking of the way a virus spreads, the identi cation of people that need help in case of a disaster (e.g., an earthquake), or just for the better understanding of the dynamics of a major event (e.g., a concert). However, only a tiny percentage of the twitter posts are geotagged, which restricts the applicability of location-based applications. In this work, we extend our framework for geolocating tweets that are not geotagged, and describe a general solution for estimating the city and neighborhood in the city, from which a post was generated. In addition, we study the speci c problem of geolocalising tweets deriving from targeted locations of interest (i.e., cities and neighborhoods in these cities), and present the visualizations of the prototype dashboard application we have developed, which can help end-users and large-scale event organizers to better plan and manage their activities. The experimental evaluation with real data demonstrates the e ciency and e ectiveness of our approach.
[Motivation:] Events that happen around us a ect our lives to di erent degrees.
The e ects of an event on a community vary depending on the type of the event
and its dynamics. For example, tra c jams a ect the way we move, football
matches and concerts may a ect the normal pace of life in the area of the venue
for a short period of time, while earthquakes and diseases are unpredicted events,
which could cause signi cant problems that have to be addressed fast. Many
entities, public and private, are interested in analyzing the e ects of such events,
in order to better understand and react to them, and lead to a better quality of
life. For example, the identi cation of lack of clean water at a place would lead
the water providers to take special care for resolving the problem. Even though
this would be a manual, labour-intensive, and time-consuming process in the
past (e.g., consider the 1854 cholera outbreak in London [
People tend to share their experiences, especially those a ecting their lives
(or feelings). Social networks, such as Twitter [
It is interesting to note that several of these applications depend on the
knowledge of the user location at the time of the posting. This knowledge is
necessary for applications that target to characterize an urban landscape, or to
optimize urban planning [
Evidently, the availability of geotagged data, determines not only the
possibility to use such applications, but also their quality-performance characteristics:
the more geotagged data posts are available, the better the quality of the results
will be (more precisely: the higher the probability for being able to produce
better quality results). Nevertheless, the availability of geotagged data is rather
limited. In Twitter, which is the focus of our study, the number of geotagged
tweets is a mere 1.5-3% of the total number of tweets [
[Proposed Approach and Contributions:] In this study, we address this
problem by extending our framework [
In contrast, in our previous work [
In this work, we extend our previous solution, and describe a general technique for estimating the location from which a post was generated using a twostage process: we rst determine the city, and then the neighborhood in the city, by building content-based models and analyzing the volume of posts over time, independently for each one of these two levels. Using this set up, we are able to e ectively predict the location of a post form the Twitter stream, when the only input we have is the actual content of the post and its timestamp.
In addition, we study the speci c problem of geolocalising tweets deriving from targeted locations of interest, that is, neighborhoods of a particular cultural, social, or touristic importance (e.g., the Vatican in Rome). Our experiments show that we can reuse our technique for this case, as well, by adjusting its operation to this context, where a small number of popular keywords mentioned in the posts characterize the location.
Finally, we present the visualizations of the prototype dashboard application we have developed, which can help end-users and large-scale event organizers to better plan and manage their activities. These interactive visualizations include heatmaps for the volume of (geotagged and geolocalised) tweets, where the user can zoom at di erent levels of granularity, ranging from a country, down to a city neighborhood, for which the user can also explore the relevant keywords. Furthermore, we provide visualizations that illustrate in a comprehensive manner the changes in the volume of posts at di erent locations over time.
[Paper Organization:] The rest of the document is organized as follows. In Section 2 we present the related work. Section 3 formalizes the problem, and Section 4 describes our solution. We present our experimental evaluation in Section 5, and our prototype dashboard implementation in Section 6. Finally, we conclude in Section 7.
Several works have studied the problem of geotagged tweet analysis. Balduini et
al. [
The problem of using tweets in order to identify the location of a user, or
the place that an event took place has been studied in the past. The \who,
where, what, when" attributes extracted from a user's pro le can be used to
create spatio-temporal pro les of users, and ultimately lead to identi cation
of mobility patterns [
The similarity between user pro les and location pro les has also been used
in [
Two recent approaches that target to geotag unique tweets are presented in
[
Two studies that target to geotag tweets are presented in [
Even though some of these studies are closely related to our work (e.g., [
The problem we address in this work is the estimation of the geographic location of individual, non-geotagged posts in social networks.
Problem 1: Given a set of geotagged posts Ptlj1 ; :::; Ptlji , t1 tj t2, where li is the location the post was generated from and tj is the time interval during which the post was generated at, and a set of individual non-geotagged posts Q1tq ; :::; Qntq , t1 tq t2, we want to identify the location l from which the post Qitq (1 i n) was generated.
The timestamps t1 and t2 represent the start and end times, respectively, of the time interval we are interested in.
In the context of this work, we concentrate on two-level ne-grained location predictions: we wish to initially estimate the coarse-grained location of a post (which is usually as big as a city) and afterwards to estimate the location at a much ner-grained level such as a city neighborhood (which is usually much smaller than a postal zipcode). Furthermore, we focus on twitter posts, whose particular characteristics are the very small size (i.e., up to 140 characters long), and the heavy use of abbreviations and jargon language. 4
In this section, we describe our solution to the problem of ne-grained geolocalisation of non-geotagged tweets. Our method is based on the creation of vectors describing the Twitter activity, in terms of important keywords, for each geolocation we have data from, and for the period of time we are interested in.
In this paper we modify the set up of our method, extending this work so
that we can use tweets from the entire stream of a social network (refer to
Algorithm 1). This is an extension of our previous work (details can be found in [
[Extraction of Important Keywords and Creation of Location
Vector:] We initialize our method by de ning the Coarse-Grained Locations (CGL)
that we target to identify posts from (i.e., cities), and the time intervals we are
interested in. Afterwards, we get the geotagged posts generated from the CGLs
in our prede ned time intervals, and we group them according to the CGL they
were generated from. We then follow exactly the same steps as described in [
Having created the CGL vectors, we repeat the procedure for the FineGrained Locations (F GL), where each F GL is a subset of a CGL (e.g., a city Algorithm 1 Tweet Geotagging Algorithm INPUT: A training set of timestamped and geotagged tweets, a timestamped querytweet (Qt) that is not geotagged, a set of prede ned coarse-grained locations (CGL) and a set of prede ned ne-grained location (F GL) where CGL F GL OUTPUT: The most eligible candidate location. 1: kwV ectorQt create vector of Qt keywords and their weights . process non-geotagged tweet Qt 2: for all i 2 fCGLg do . process training dataset, for all coarse-grained locations 3: for all t 2 ftime intervalsg do . and for all time intervals 4: Docit all tweets in location i at time interval t 5: kwV ectorit create vector of Docit keywords and their weights 6: similarityi V ectorSim(kwV ectorQt ; kwV ectorit ) 7: if similarityi > 0 then 8: Add CGL to candidate coarse grained locations (CandCGL) 9: for all j 2 CandCGL do 10: for all F GLi 2 j do 11: for all t 2 ftime intervalsg do . and for all time intervals 12: DocF GLit all tweets in location F GLi at time interval t 13: kwV ectorF GLit create vector of DocF GLit keywords and their weights similarityF GLi V ectorSim(kwV ectorQt ; kwV ectorF GLit ) T woLevelSimilarityF GLi similarityj similarityF GLi
argmaxi2F GLsfP robCalc(T woLevelSimilarity) . identify location of 14: 15: 16: location
tweet Qt 17: return location neighborhood), extracting the keyword vectors that describe the activity for the level of the F GLs.
[Similarity Calculation and Best Match Extraction:] We then
calculate the similarity between the keyword vector of Q and the keyword vector
of each one of the CGLs, using the function V ectorSim (as described in [
Having extracted the most eligible CGLs, we proceed to the next stage and check all the F GL that are subsets of an eligible CGL (once again using the V ectorSim function). Although we have already ltered the CGLs with low similarity, it is possible to encounter low-similarity scores at the level of F GL, as well. In order to avoid this, we use an additional data-driven threshold, and we extract the F GLs that exceed the threshold of the ne-grained granularity. We then combine the F GL similarity with the CGL similarity, multiplying them and getting a unique value for each eligible F GL, which we normalize and convert Algorithm 2 Probability Calculation 1: procedure ProbCalc(similarities between Qt and candidate geolocations (Geolocs)) 2: for all i 2 Geolocs do . Get the probability distribution 3: P robit;Qt PSimSiitm;QQtt 4: SortDescendingP robit;Qt 5: return Geolocs and their P robit;Qt into a probability. At the end, the algorithm returns the geolocation with the highest probability (refer to Algorithm 2).
[Similarity Based on Correlation of Activity Time Series:] As
established in our previous study [
The correlation factor is a parameter that allows us to exploit the timeevolution behavior of a location: we initially extract the correlation between \global" and \local" locations, and afterwards we multiply it to the similarity extracted between the keywordV ectorlocal and the keywordV ectorQ. in our current study, we employ this feature, but restrict it only to F GLs, since (as we explained earlier) we are now dealing with posts that may not correspond to any of the candidate locations. 4.1
As we have already mentioned, the set up of this method allows us to identify non-geotagged tweets that are coming from the full stream of a social network. As a result, posts irrelevant to our candidate locations could still share stopwords, leading to a (small) similarity to some location. In order to lter out these cases, we use thresholds on the similarity, both for the CGLs and the F GLs.
The distribution of the keywords among the candidate locations is di erent depending on the time intervals we check. Therefore, the signi cant keywords are going to have di erent weights for each time interval. For example, during the night we have a few posts, leading to the creation of small dictionaries, where matching one of the stopwords in these dictionaries would lead to high similarity between the Q tweet and the candidate location. In this case, the threshold should be set high. This is not true when we consider the dictionaries created during the day.
In order to automatically set a dynamic threshold, we use a small training dataset (in our case 1 day), keeping the similarities between each Q tweet and the location that it corresponds to. We initiate our threshold by setting it to 0. Then, we identify the tweets that are correctly matched to a location, and we record their similarity. At the end, we compute the mean of all the similarities, giving us the threshold extracted from the rst day and for the speci c time intervals. In order to set up the threshold for these time intervals for the following day, we use the mean of the thresholds used in all previous days. As a result, the threshold for a given day and time interval is computed as the mean of the threshold means of the previous days for the same time interval.
Following the procedure described above, we dynamically update the threshold: the thresholds are data driven, and the method is parameter-free. 5 5.1
We performed the experiments on a PC with an Intel Core i7-2600 CPU @ 3.40GHz x 8 processor, running Ubuntu 16.04.2 LTS with 8GB RAM. We implemented our algorithms in Python 2.7, and used the Geoplotlib toolbox for the visualizations.
[Datasets:] We use a real dataset containing geotagged posts from Twitter, generated in Italy between June 1 and June 20, 2016. The coarse-grained locations that we focus on are the 7 Italian cities with the highest activity, namely Rome, Milan, Florence, Venice, Naples, Turin and Bologna. The total number of tweets is 218,572: 23,566 originated from Rome, 18,824 from Milan, 7,840 from Florence, 4,628 from Venice, 4,071 from Naples, 3,719 from Turin, 2,624 from Bologna and 153,300 from the rest of the country. As ne-grained locations, we consider city neighborhoods represented by 1km-side squares. For targeted locations, we have selected the Vatican (a 1.3km-side square) from Rome, and San Siro Stadium (a 0.8km-side square) from Milan. The time windows we use have a duration of 4 hours (which can e ectively capture an important event, as well as the start and the aftermath of this event), while also keeping the detailed aggregated information for every 15min time interval.
Instead of using tumbling windows as done in [
[Algorithms:] We experimentally evaluate the two-level algorithm we
described in Section 4. In order to initialize our thresholds, we run our method
on the rst day, but exclude these results from the evaluation. In all our
experiments, we randomly divided the dataset in 80%training and 20% testing,
repeated each experiment 5 times, and reported the mean values in the results.
We compare the results of our two-level algorithm with the state-of-the-art QL
and KL algorithms [
[Evaluation Measures:] We study the e ectiveness of our approach using the precision and recall measures: P recision = cggTTwweeeettss and Recall = cagTTwweeeettss , where cgT weets is the number of the correctly geolocalised tweets, gT weets is the number of tweets we geolocalised, and aT weets is the number of all tweets that are originally deriving from our candidate locations. We also report the balanced F1 measure, F 1 = 2 PPrreecciissiioonn+RReeccaallll . 5.2
In the rst set of experiments, we study the performance of our method when we vary the number of CGLs (i.e., cities) between 1 and 7. As estimated location, we only consider the rst answer given by our algorithm (i.e., @Top1). We note that the random algorithm had precision less than 0.024% and recall less than 0.12%, with the highest values occurring when using 1 city.
In Figures 1a-1b, we illustrate the precision and recall when we use 7 CGLs. The results for 1-6 CGLs are very similar, and omitted for brevity. The F1 for the cases of 1 and 7 CGLs are compared in Figures 1c and 1d, respectively. Using our approach, we achieve a precision of up to 89%, and a recall of up to 17%, while the best F1 was 26%. The best precision is achieved when using 60% of the keywords and a threshold of +20%, while the best recall and F1 for 70% of the keywords and \no threshold".
For the comparison to the state-of-the-art presented in Figure 2, we use the version of our method with threshold +20%. As depicted in the plots, our method achieves up to 80% precision and 23% recall, while KL only achieves up to 13% precision and 20% recall.
We note that as we increase the number of CGLs considered, we would expect to see a reduction in the precision and recall values, as a result of the increased search space. When looking at all the detailed results though, we do not observe this. On the contrary, precision slightly increases as we add CGLs, demonstrating the robustness of our approach. 5.3
We now evaluate the performance of the proposed approach for targeted locations of interest. The results for the Vatican and San Siro locations are presented in Figures 3a-3b, and Figures 3c-3c, respectively.
The precision for Vatican reaches a maximum of 68% when using either 10%, or 20% of the keywords and \no threshold", while recall reaches 84% when using 100% of the keywords and \no threshold". Similarly, San Siro achieves a precision of 49%, and a recall of 54%, for 10%, or 20% of the keywords, and for 100% of the keywords, respectively, and \no threshold". These numbers correspond to a pretty high performance, especially when taking into account the very high recall values.
We note that in both locations, the precision and recall values are exactly the same when using 10% and 20% of the keywords, while the precision reduces suddenly after that. A close look at the dictionaries of the two locations revealed that the most important keywords are the names of the locations. The small (a) Precision for 7 CGLs (b) Recall for 7 CGLs (c) F1 for 1 CGL (Milan) (d) F1 for 7 CGLs dictionary size employed (when using 10-20% of the keywords) is then occupied by these keywords. As the dictionary size increases, stopwords and noise are inserted, which have a negative impact on precision. 6
We now provide a brief description and sample screenshots from our prototype dashboard implementation, which uses our tweet geolocalisation solution in order to help users visualize and better understand geolocalized Twitter activity.
Figure 4a depicts the geotagged tweets that are generated from Italy in the form of a heat-map (the more red the color of a location, the more tweets originate from that location). Due to the bounding box used for extracting the tweets, we have tweets from places outside of Italy, as well. Nevertheless, this does not pose a problem, since our approach aims to geotag any tweet in the public stream, and is language independent. As we can observe from the heatmap, the Twitter activity is heavy in the big cities (e.g., Milan and Rome), and low in the country side.
The user can then use the map to zoom in, for example on the capital of Italy, Rome. In this case, we observe that the activity around important locations, such as the Vatican and the Colosseum, is much higher than the activity at the suburbs of the city (refer to Figure 4b).
If the user continues to zoom in, they arrive at the detailed city view, illustrated in Figure 5a. In this view, we can see the square grid structure, superimposed with a heat-map that represents the distribution of the geolocalised tweets (the more tweets a location has, the more dense the color is). In addition, this view allows the user to examine the topic models computed for each one of the neighborhoods: when the mouse hovers over a square, a window pops up that lists the most important keywords for that location.
Another useful analysis tool is the study of how the volume of tweets changes over time in di erent neighborhoods of a given city. This is depicted in Figure 5b, where we report the relative percentage increase(/decrease) of the volume of tweets in the current time interval, when compared to the previous time interval (i.e., (#NewT weets #OldT weets) ). In case we have data in a square, we put the #OldT weets percentage of the di erence with the previous window, otherwise the square appears without any number. This view can quickly reveal neighborhoods, where the twitter activity is rapidly increasing(/decreasing), signifying, for example, an aggregation of people in a speci c location of the city.
More speci cally, in Figure 5b, we present how the activity changes at Milan while getting closer to the beginning of the soccer game in the San Siro Stadium. As the gure shows, the square in which San Siro Stadium is located and it's (a) Precision and Recall (b) F1 Fig. 2: Precision, Recall and F1 Comparison for 7 CGRs (@Top1) (a) Precision (b) Recall (c) Precision (d) Recall Fig. 3: (Top) Vatican (1.3km-side square / @Top1). (Bottom) San Siro (0.8kmside square / @Top1). neighboring square have an increase of the activity, while the majority of the squares in Milan have a decrease. After further analyzing the activity of the squares, we identi ed that the square in which the San Siro Stadium is located has a constant increase of the activity up to 1 hour before the beginning of the match, while there is no other square with a constant increase. This representation allows us to depict the e ect that such an event has on the area, and provides local authorities useful information that can be used to o er better services to the citizens (e.g., better manage the tra c ows).
Finally, the user can focus on a speci c location in the city, and view the volume of tweets over time for that location. Figure 6 illustrates the tweet activity time series5 for the Vatican and San Siro locations (for the time interval of 5 The labels of the peaks have been inserted manually, but could also been done automatically by analyzing the news streams. May 26, 17.00 to June 1, 17.00, during which we had some important events). These graphs show that these two locations have very di erent characteristics: Vatican exhibits a relatively low, yet stable activity stream; on the other hand, San Siro has almost no activity for a large part of the time interval, but includes activity bursts. Nevertheless, it is interesting to note that, as reported in Figure 3, our algorithm for tweet geolocalisation performs equally well for both targeted locations.
(a) Italy (b) Rome
(a) Volume of Geotagged Tweets of (b) Di erence in the Activity of squares Rome and Representative Keywords around San Siro after one slide (May for Vatican square (June 2, 10:00- 28, [15:00-19:00] to [15:30-19:30]) 14:00).
Fig. 5: Geotagged Tweets, Keywords and Activity Di erence (a) Vatican
(b) SanSiro Stadium In this work, we present a framework that allows us to geolocalise non-geotagged tweets. Our two-level framework allows the estimation of the location from which a post was generated, by exploiting the similarities in the content between this post and a set of geotagged tweets. Contrary to previous approaches, our framework provides geolocation estimates at a ne grain, thus, supporting a range of applications that require this detailed knowledge and could be used for social good. The experimental evaluation with real data demonstrates the e ectiveness of our approach, regardless of the number (as small as 1) and size (as small as a 700m-side square) of the targeted locations. In order to render the results of our methods easier to understand, we developed a prototype dashboard that provides interactive visualizations, assisting end-users and large-scale event organizers to better plan and manage their activities.
This work was supported by a fellowship from Telecom Italia.