Proceedings of the 2015 Workshop on Vision and Language (VL’15), pages 51–59,Lisbon, Portugal, 18 September 2015. c©2015 Association for Computational Linguistics.

Understanding Urban Land Use through the Visualization of Points ofInterest

Evgheni PolisciucCISUC

Departement ofInformatics EngineeringUniversity of Coimbra

Coimbra, Portugalevgheni@dei.uc.pt

Ana AlvesCISUC, DEI

University of Coimbra& Polytechnic Institute

of CoimbraCoimbra, Portugalana@dei.uc.pt

Penousal MachadoCISUC

Departement ofInformatics EngineeringUniversity of Coimbra

Coimbra, Portugalmachado@dei.uc.pt

Abstract

Semantic data regarding points of interestin urban areas are hard to visualize. Due tothe high number of points and categoriesthey belong, as well as the associated tex-tual information, maps become heavilycluttered and hard to read. Using tradi-tional visualization techniques (e.g. dotdistribution maps, typographic maps) par-tially solve this problem. Although, thesetechniques address different issues of theproblem, their combination is hard andtypically results in an efficient visualiza-tion. In our approach, we present a methodto represent clusters of points of interest asshapes, which is based on vacuum pack-age metaphor. The calculated shapes char-acterize sets of points and allow their useas containers for textual information. Ad-ditionally, we present a strategy for plac-ing text onto polygons. The suggestedmethod can be used in interactive visualexploration of semantic data distributed inspace, and for creating maps with simi-lar characteristics of dot distribution maps,but using shapes instead of points.

1 Introduction

Understanding the urban land use is one of thecentral pillars of urban planning and management.Traditionally this analysis relies on census surveyshaving limitations in terms of spatial and temporalscale. However, with the advent, and wide deploy-ment of pervasive computing devices (e.g. cellphones, GPS devices, smart cards and digital cam-eras) some of these limitations may be overcome.For instance, collecting and analyzing informationof how people use urban space may be done dy-namically and in more precise way.

By using services of modern web platforms(e.g. Facebook, Foursquare, etc.), a user leaves

“digital footprints”. These are precise data interms of temporal (when) and spatial locations(where), and in general, can be captured withouthuman intervention. The information about hu-man activity (what) if not explicitly introduced byhumans may be inferred by other ways. One ofwhich is about to retrieve the information aboutvisited place. This place, denominated Point of In-terest (POI), offers a range of services and has spe-cial utility. Such information is not always avail-able. Hence, it is necessary to enrich semanticallythe information about the visited places, in order tounderstand what was done there. Collecting infor-mation of how people use urban space has becomea very important task on creating the city imagefrom the perspective of its inhabitants, since placesare often associated by meaning, i.e. relationshipbetween people and places.

Most smart devices integrate contextual pro-cessing. However, it is difficult to enable context-awareness without semantic information. Al-though, semantic information has been availablefor years, the Internet, in most cases, abandonssuch information. In a recent work Alves (2012)presented various perspectives on semantic enrich-ment of places and extraction of such informationfrom the Internet.

That said, there is a necessity of proper visual-ization that depicts large amounts of point-baseddata along with textual information. Geovisual-ization field provides techniques to visualize geo-referenced data, known as thematic maps. Oneof the well known techniques to represent point-based data is a dot distribution map. However, thiskind of maps is limited to representation of pointson the map, additionally using color to distinguishpoints that belong to different groups. On theother hand, typographic maps are used to representtextual information on the map regarding naturaland artificial features of urban space (e.g. streetnames, rivers, places, etc.). But, in order to visu-

51

alize both textual and point-based information onecannot simply overlay two maps. In this case thevisualization becomes highly cluttered and illegi-ble. Moreover, it would be difficult to reveal spa-tial patterns in such hard-overlapped maps. There-fore, from these observations we propose a methodto represent this kind of information in a visual-ization with low degree of visual clutter retainingthe possibility to both reveal high-level informa-tion and detailed exploration of the map.

Our approach consists in creating visual ele-ments that convey spatial distribution of POIs ofsame type (a cluster), as well as the distributionof clusters in urban area. More precisely, our al-gorithm generates a shape for each group of POIsrevealing its unique visual form in regard to theirgeographic distribution. Additionally, textual in-formation – clusters tags and POI names – aredrawn using different typeface weights and scaledaccording to the relevance of each cluster.

With that said, in this paper we present a methodfor visualizing clusters of POIs and the associatedsemantic information. The dataset is detailed insection 3. The shape of each cluster is calculatedusing a vacuum package metaphor (see details insection 4). Additionally, this paper presents aninteractive web-based application that allows ex-ploring the data with varying degree of details (seesection 5).

2 Background and Related Work

Our approach touches on diverse methods andtechniques of visualization of spatial information.In this work we consider dot distribution maps.This type of maps are especially efficient in visual-ization of distribution and densities of point-baseddata. Regarding the visualization of textual infor-mation our approach relies on typographic maps.This particular type of subjective maps efficientlycommunicates textual information prioritizing ty-pographic hierarchy depending on the relevance ofinformation.

Dot distribution maps, often referred to asdensity map, they represent spatial distribution ofgeo-referenced data using basic graphical element– a point (Slocum, 2009). Each point on the mapis used to represent either one datum with knowngeo-location, or aggregation of values. Addition-ally, dot distribution maps are used to depict densi-ties in corresponding geographic areas, rather thanspecific locations.

A historical example of the use of a dot dis-tribution map is the disease map produced byJohn Snow (Tufte and Graves-Morris, 1983). Thismap depicts the distribution of cholera in London.Deaths are represented by dots and eleven waterpumps are represented by crosses. The observa-tion led Snow to discover that cholera occurred inthe areas near the Broad Street water pump. Thismap helped understand the issue of the cholera byrevealing disease patterns in spatial context.

One of a more recent example of density map isthe Racial Dot Map by Cable (2013). This visu-alization depicts geographical distribution, popu-lation density and racial diversity of people livingin USA. Each dot represents one individual personat smaller zoom levels and aggregation of dots atnational or regional levels. The color encodes raceand ethnicity of inhabitants.

Typographic maps may be seen as an “artis-tic” representation of textual information, ratherthan an accurate mapping of spatial data. Often,the information being represented by these mapsis a description of the relationship of the place andits meaning, which depends of many human, cul-tural, political, social or historical factors. There-fore, these kind of maps are considered subjectivemaps (Chen, 2011).

In typographic maps, as the name indicates,textual information is represented using typogra-phy. For instance, the maps drawn by Paula Scherare mainly typographical, representing the world,its continents, countries, islands, etc. throughtypography (Scher, 1990 2010). Likewise, themaps produced by Axis Maps, depict the infor-mation about locations and space using text (AxisMaps, nd). Moreover, the geometry of each wordis curved along a path, mimicking the shape ofthe object being represented (e.g. streets, parks,rivers, etc.). This typographic maps were com-posed with auxiliary of software-based tools (e.g.Adobe Illustrator) and represent information us-ing digital typography. Finally, the graphical el-ements are placed over OpenStreetMap. Theseworks, the maps by Scher and axisMaps, are goodexamples of intelligent usage of typographical hi-erarchy, which makes these maps efficient in thecommunication of subjective and imprecise infor-mation, even with high degree of visual overload.

A more recent research presents a methodfor automatic construction of typographic mapsby merging textual information with spatial data

52

(Afzal et al., 2012). Given a vector map the al-gorithm places textual labels in space along thepolylines and polygons in accordance with definedvisual attributes and constrains. Additionally, theauthors describe a method to represent regions astext by filling its interior and repeating the text asnecessary. Likewise, our approach uses principlesof this technique to align textual labels to a path.

Finally, the work of Cranshaw et al. (2012) istightly related to our approach, especially in whatconcerns portraying a city using methods to visu-alize point-based data and their clusters. The au-thors introduce a method that consists of a clus-tering model for mapping a city regarding collec-tive behaviors of its inhabitants and further visu-alization on the map. This map depicts dynam-ics, structure and portrayal of a city using clus-ters, so called Livehoods, of geospatial data fromFoursquare check-ins. Given geospatial socialdata generated by hundreds of thousands of peo-ple the visualization represents distinct areas of thecity regarding activity patterns. The resulting ag-gregated clusters of check-ins represent so calledmental map of the city, the vision of urban spacefrom the perspective of its inhabitants. This en-ables the study of the structure and composition ofa city based on social media its residents generate.

3 Data Description and DesignRequirements

Our dataset consists of points of interest (POIs)from the greater metropolitan area of Boston,Massachusetts, USA. POIs contain associated se-mantic information and are aggregated in mean-ingful groups (e.g. restaurants, colleges, indus-try, etc.). More precisely, POIs were taggedwith semantic information retrieved from diverseweb sources (e.g. Foursquare, Upcoming Yahoo,etc.) (Oliveirinha et al., 2010), and aggregatedin clusters using methods proposed by Alves etal. (2011). The dataset comprises 751 clusters ofPOIs with the following attributes: tag and id ofeach cluster, geographic coordinates of their cen-troids, and relevance of a cluster. Additionally,each POI in the dataset is characterized by ge-ographic location (latitude and longitude), nameand id. Ultimately, the data types are categorical– POI names and cluster tags – and quantitative –relative relevance of each cluster.

In order to guide the design of our visualization,we established the following design requirements,

that define the boundaries for the project:

• The visualization should create a digital layerof urban space.

• It should use a simple and clear visual lan-guage, establishing a strong relationship be-tween urban space and POIs.

• In order to reflect geographic nature of datathe information should be visualized on amap.

• It should be interactive and run in real-time,supporting the process of data explorationand high-level information acquisition.

• The interactive application should follow theso called Visual Information-Seeking Mantra,introduced by Shneiderman (1996), whichconsists of overview first, zoom and filter,then details-on-demand.

• Finally, the visualization should be easy-to-understand by a general user with no ana-lytic background, therefore presenting a goodbalance between aesthetics and functionality,without visual overload of display.

4 Representation of POI Clusters

This section covers the process for determining theshapes that describe POI clusters. More precisely,first the concept of vacuum package metaphor isintroduced. Then we proceed with the descriptionof an algorithm for polygon calculation given a setof points. Then, we present a method for smooth-ing the corners of generated polygons. Finally, wediscuss the strategy for using typeface weight asvisual variable and text placing.

4.1 Concept

In order to understand the distribution of POIs inspace and within the corresponding clusters weplotted them using a dot distribution map (see Fig-ure 1). In this visualization each POI is depictedby a point; The category it belongs to is repre-sented with a color. The observation of this visu-alization led us to the conclusion that each clus-ter has its unique and recognizable shape. Forinstance, the same happens with the shapes forcountries and continents of the world, to whichdiverse meanings and symbolisms are associated.

53

Figure 1: Direct representation of data – names of categories and POIs represented as colored points.Points of equal color belong to the same cluster.

We found this idea particularly interesting and im-plemented a method that allows us to character-ize sets of POIs through their shapes. This ap-proach allows us to eliminate color from represen-tation reducing visual overload. Ultimately, due toamount and diversity of categorical data the use ofcolor is inefficient.

The shape that characterizes a given set ofpoints is known as convex hull, i.e. the polygonthat encompasses a set of points. This is, the min-imal convex set of points containing the entire set.One of the many algorithms to compute convexhull was proposed by Andrew (1979), which isbased on the rubber band metaphor. The con-vex hull can be visualized as a rubber band thatis stretched so that it surrounds all the points andthen released enclosing all the points.

For certain applications the convex hull does notrepresent well the boundaries of a set of points.For instance, a convex hull for a set of points thatform a C letter would have a shape close to ellipse.In other words, the region that is defined by convexhull does not represent the region that is formed bythe points.

This problem has already been addressed bymany researches, and is know as computation ofnon-convex or concave hull (see e.g. Moreira andSantos (2007) and Duckham et al. (2008)). One

of the methods to compute shape of a set of pointswas introduced by Edelsbrunner et al. (1983), andis know as two-dimensional alpha-shapes. Asmentioned by Edelsbrunner, an alpha shape canbe imagined as a huge mass of ice-cream con-taining pieces of chocolate; then using a sphere-formed ice-cream spoon we carve all the parts ofice-cream without bumping into chocolate pieces;if we now straighten all the ”round” faces we willget an intuitive description of what is called thealpha-shape. Although, the alpha-shape is a stan-dardized formal description of a set of points thismethod assumes multi-polygon reconstruction ofthe geometry, and often produces polygons thatcontain holes, which is not desirable in our visual-ization.

We address this problem by combining theprinciples of alpha-shape and vacuum packingmetaphor. This metaphor can approximately beseen as following: imagine a plastic bag contain-ing a set of points; then the oxygen is removed cre-ating vacuum inside the bag; when the bag is com-pletely shrunk it become tightly fitted to its con-tent. This metaphor is intuitive description of ourapproach. The following section gives a formaldescription of the algorithm to compute a shape ofa set of points heterogeneously distributed in geo-graphic space.

54

4.2 Algorithm

This section describes an algorithm for the calcu-lation of a concave hull based on the vacuum pack-age metaphor. The calculation of a polygon is aniterative process with the maximum number of it-erations defined by an user. The process passesthrough the calculation of convex hull, which de-fines an initial set of edges. Each edge is char-acterized by starting and ending points in orderedarray, and by stretchiness, which models a behav-ior of an elastic band. Finally, the shape of eachset of points is calculated independently.

Considering the set of points S to be our input,the algorithm proceeds as follows:

1. Let L be an initially empty list that will con-tain the points that define the polygon.

2. Calculate the convex hull, and store the set ofpoints in L in clock-wise order.

3. For each iteration and for each edge – i.e., foreach pair of consecutive points in L, whichwe designate by A and B:

• If the length of the edge is bigger thanpredefined minimum length, then con-tinue to the next step. Otherwise, skip.• The edge AB is divided by half at

the center, defining an isosceles triangle4ABC, where A and B are the startingand ending points in L, respectively, andC is the central point.• C is pushed inside the polygon by a

force vector ~f , which is perpendicularto the edge AB.• The magnitude of the force ~f varies pro-

portionally to the stretchiness of ABedge, which is a function of its length,and the distance of C from its originallocation, say M . i.e. shorter edges havesmaller ~f .• If one of the points, say P , in the set S

is inside the triangle, then the P is ap-pended to the L in the order AP andPB, consequently, creating two newedges.• The process is repeated until the maxi-

mum number of iterations is reached orall the edges have their lengths smallerthan defined minimum length.

Determining if a point P is inside the triangle4ABC is done by calculating cross products ofvectors ~AP× ~AB, ~BP× ~BC, and ~CP× ~CA. If allthe values are negative, then the point is inside thetriangle. Otherwise, the point is outside. Finally,at a simulation instance there might be two pointsinside the triangle. In this case considered only theone that is closest to the M point – middle pointthat divides AB by half.

Figure 2 displays two shapes of clusters thatwere calculated by the algorithm given two sets ofpoints from our dataset. Also, this figure schemat-ically illustrates the described algorithm at a sim-ulation instance – the points that compose a poly-gon are marked with circles, the edges are repre-sented with black line, and the lines that makeupthe triangles are painted in red. As can be ob-served in Figure 2, image on the right, even com-plex shapes are well defined.

Figure 2: Calculated shape of clusters for ”Trad-ing”, image on the left, and ”Seinfeld” categories,image on the right. Circles represent points thatcompose a hull, with the arrows inside that indi-cate the order of points.

4.3 Visual Refinement and Label Placing

Having calculated all the polygons we proceed tosmooth their corners, which gives an organic rep-resentation of a shape. Also, this facilitates theprocess of placing textual labels onto polygons.

The problem of corner smoothing can be di-vided in two parts – round the corners that makeinterior angles smaller and greater than 180◦ (forthe purpose of simplicity we label them as S and Gcorners, respectively). In order to create a smoothpolygon there should be enough free space allo-cated to append additional points that compose a

55

new polygon. This is done by translating perpen-dicularly each edge to a certain distance, say d,placing them outside the original polygon. Thenwe proceed by calculating the S corners, simplyconnecting two consecutive translated segmentswith an arc that is centered at the corner pointand with the radius equal to d. The arc is thenfragmented with small segments, the number ofwhich is dictated by defined minimum length. TheG corners, on the other hand, are computed us-ing Bezier equation (Farin et al., 2002). The twocontrol points have the same location, which is thepoint of intersection of the two edges that makeupthe corner. The end points are the middle pointof each of two translated edges. Finally, the curveis partitioned with small segments. The Figure 3is a schematic representation of the smoothing ofpolygons.

Figure 3: Schematic construction of round cor-ners.

The second part of visual refinement consists ofplacing textual information onto polygons. As wasmentioned earlier, our data consist of two types oftext data – most relevant tag for each cluster, andnames of each individual POI. In order to repre-sent this data two different strategies were used.In the first case, the cluster tags were positionedon center of the polygon being represented usingtypeface weights to represent their weights. In thesecond case, the names of POIs were placed ontothe polygon and curved along the contour of theshape.

As mentioned above, we used typographicweight as a visual variable to represent relevanceof each category of POIs. According to Lupton(2014), in typography, all typefaces are organizedinto families . Within a family typefaces are di-vided and ordered according their weights (e.g.regular, bold, etc.). In modern typography thereare families that contain up to nine weights classi-

fied as fallowing – thin, extra light, light, regular,medium, semi-bold, bold, black and heavy. So,we used these to encode the relevance of each cat-egory, by dividing all values in eight ranges andassigning each range to a typographic weight. Dueto ordered nature of data the typographic weightswere assigned starting with thin up to heavy type-faces (see Figure 4).

Figure 4: Close-up of an area with categories thathave different relevance encoded with typographicweight and font size.

For the second type of textual information weused an approach inspired in typographic maps.The name of POIs are placed onto polygon andthe words are curved along with the shape. One ofthe problems of curved words is the fact that theyvisually distort the word when placed on the pathjunctions. This is, the words are visually breakingapart creating discontinuous reading. One of thesolutions to address this problem is to distort thecharacters, like in maps by Paula Scher. However,in digital typography these manipulations are un-desirable and are considered a bad practice (Lup-ton, 2014). So, our solution consisted in: draw-ing the letters perpendicularly to the path they areplaced onto; when a letter appears on a junctionof two segments we use a weighted angle depend-ing on the percentage of occupied space on eachof segments. In other words, the imaginable rect-angle that holds a character always keeps its basecorners on top of each segment (see Figure 5). Fi-nally, the tracking – space between characters – isincreased, when the letters are placed on G cor-ners, and decreased, when the letters are placed onS corners. This diminishes the visual discontinu-ities in reading.

Finally, all the methods were combined and theresult is displayed on Figure 6. As can be observed

56

AFigure 5: Placing characters onto segment andsegment joint.

the visualization become less cluttered in compar-ison with dot map representation. The clusters ofPOIs are characterized by an organic shape, whichfacilitates the placement and continuous readingof textual information. Text labels, on the otherhand, try to mimic the contour without substantialvisual distortion. Also, it is easy to understand thatthis approach is less efficient when small clustersare considered, due to limited space to display allthe textual information.

5 Application and Limitations

This section presents an application of the de-scribed techniques combined into one visualiza-tion model. First, an interactive web-based appli-cation, which uses the described methods appliedon another dataset with a similar nature of infor-mation, is discussed. Then, we enumerate the lim-itations of presented approach.

The interactive web-based application fol-lows the principles of Visual Seeking Mantra –overview first, zoom and filter, then details-on-demand. In the first screen the user can find a gen-eral view of the map. In this view the visualizationdepicts only the shapes of clusters, which gives afirst impression about the data and its distributionin space. It is important to note that in the seconddataset POIs have multiple associated tags, i.e. aPOI may belong to different categories. Conse-quently, POI clusters may overlap, which meansthat overlapping areas provide multiple services.For instance the area of restaurants might coin-cide with the area of shopping. That said, the usercan easily identify these cases in general view (seeFigure 7, image on the left).

Filtering and zoom-and-pan are also importantfunctionalities of the application. Using the panelon the left the user can select individual categoriesto display and filter the visualization by averageweight of the relevance, by the number of POIsin cluster, among others. Selecting the categoriesalso reveals their names and places them as de-

scribed in previous sections, although using onlyone typographic weight. To navigate on the mapthe user can use zoom-and-pan. The visualiza-tion dynamically updates details of the shapes andpresents different levels of cluster aggregation ac-cording to zoom level (see Figure 7, image in themiddle).

Finally, the application provides additional de-tails on demand. This is done by directly select-ing clusters on the map. In this case the panel onthe left updates and displays more detailed infor-mation about the selected cluster (e.g. a list ofPOIs in the group, impact of each category thecluster belongs to, number of POIs). Additionally,the clusters that share the same category are alsohighlighted on the map, such that the distributionof clusters within similar category is revealed (seeFigure 7, image on the right).

As can be observed in the web-application thelabels are not shown, due to high amount of textualinformation, which makes the visualization runslow in a web browser. Nevertheless, this func-tionality is implemented in offline visualization.As it can be observed, there are overlapping areas.As such, the shapes are painted with transparentcolor, in order to highlight highly overlapped areason the map. Allowing the user to perceive urbanareas that provide multiple services can be easilyfound on the map. Thus, providing higher-levelinformation that would be difficult to visualize byother means.

6 Conclusion

In this article, we presented a method to representclusters of POIs along with their semantic infor-mation. This method integrates visual characteri-zation of a set of points and the methods to repre-sent textual information. Given clusters of POIsthe presented method creates a visual layer thatcharacterizes urban space in accordance with themeanings of places, which derives from the dig-ital footprints that the inhabitants leave. For thisreason, we presented a novel approach that cal-culates a concave hull of a set of points. Thismethod enables the creation of a unique integralpolygon, which is calculated using vacuum pack-age metaphor. Ultimately, each polygon character-izes a set of points with a unique organic lookingshape.

Additional textual information is added by plac-ing names of POIs on a path defined by a poly-

57

Figure 6: A zoom-in of downtown area of Boston. The final visualization combines all the presentedmethods: shapes of clusters; rounded corners; label placing.

Figure 7: Screenshots of interactive application. General view, image on the left; filter and zoom, middle;and details-on-demand, image on the right.

gon. In the proposed strategy the characters wereplaced perpendicularly to the segment they be-long, and using a weighted angle when placed onthe corners. Additionally, we introduced a visualvariable to encode the relevance of a category –the typographic weight. The data variable’s valuesare divided in equal ranges, and then each bin isassociated with a typographic weight in the orderfrom thin to heavy.

Finally, this visualization was implemented asweb-based application and applied on anotherdataset. The interactive web application gives an

overview of data at general zoom level. Then theuser can zoom-in and obtain a detailed view of thevisualization. Additionally, using the filter panelthe user can choose individual category and filterthe visualization by different parameters. Finally,more details, such as cluster impact or the list ofPOIs in cluster, are given on demand.

Acknowledgements

This work was supported by the InfoCrowdsproject - FCT-PTDC/ECM-TRA/1898/2012FCT.

58

ReferencesShehzad Afzal, Ross Maciejewski, Yun Jang, Niklas

Elmqvist, and David S Ebert. 2012. Spatialtext visualization using automatic typographic maps.IEEE Transactions on Visualization & ComputerGraphics, (12):2556–2564.

Ana O Alves, Filipe Rodrigues, and Francisco CPereira. 2011. Tagging space from information ex-traction and popularity of points of interest. In Am-bient Intelligence, pages 115–125. Springer.

Ana Cristina da Costa Oliveira Alves. 2012. Semanticenrichment of places. understanding the meaning ofpublic places from natural language texts.

Alex M Andrew. 1979. Another efficient algorithm forconvex hulls in two dimensions. Information Pro-cessing Letters, 9(5):216–219.

Axis Maps. n.d. Typographic maps. http://store.axismaps.co.uk/. Accessed: 2015-06-10.

Dustin Cable. 2013. Racial dot map. Weldon CooperCenter for Public Service, University of Virginia.

Xiaoji Chen. 2011. Seeing differently: cartographyfor subjective maps based on dynamic urban data.Ph.D. thesis, Massachusetts Institute of Technology.

Justin Cranshaw, Raz Schwartz, Jason I Hong, andNorman M Sadeh. 2012. The livehoods project:Utilizing social media to understand the dynamicsof a city. In ICWSM.

Matt Duckham, Lars Kulik, Mike Worboys, andAntony Galton. 2008. Efficient generation ofsimple polygons for characterizing the shape of aset of points in the plane. Pattern Recognition,41(10):3224–3236.

Herbert Edelsbrunner, David G Kirkpatrick, andRaimund Seidel. 1983. On the shape of a setof points in the plane. Information Theory, IEEETransactions on, 29(4):551–559.

Gerald E Farin, Josef Hoschek, and Myung-Soo Kim.2002. Handbook of computer aided geometric de-sign. Elsevier.

Ellen Lupton. 2014. Thinking with type. ChronicleBooks.

Adriano Moreira and Maribel Yasmina Santos. 2007.Concave hull: A k-nearest neighbours approach forthe computation of the region occupied by a set ofpoints.

Joao Oliveirinha, Francisco Pereira, and Ana Alves.2010. Acquiring semantic context for events fromonline resources. In Proceedings of the 3rd Interna-tional Workshop on Location and the Web, page 8.ACM.

Paula Scher. 1990–2010. Typographic maps. http://paulaschermaps.com/. Accessed: 2015-06-18.

Ben Shneiderman. 1996. The eyes have it: A taskby data type taxonomy for information visualiza-tions. In Visual Languages, 1996. Proceedings.,IEEE Symposium on, pages 336–343. IEEE.

Terry A Slocum. 2009. Thematic cartography andgeovisualization. Prentice hall.

Edward R Tufte and PR Graves-Morris. 1983. Thevisual display of quantitative information, volume 2.Graphics press Cheshire, CT.

59