April 15-18, 2009
Indianapolis, Indiana, USA

The Interpretation of Bias (and the Bias of Interpretation)

Aaron Straup Cope, Flickr


This is a story about Flickr, a popular online photo-sharing Web site, and some of the lessons learned by allowing users to add their photos to a map.

This Is A Story About Naming Things

A map is, in its primary conception, a conventionalized picture of the Earth's pattern as seen from above. (Erwin Raisz)

Every map is someone's way of getting you to look at the world his or her way. (Lucy Fellows)

Geocoding is the act of converting a named place or address into machine readable coordinates, typically a latitude and longitude using the Mercator projection. This is Raisz' "conventionalized pattern".

Reverse-geocoding is the act of taking a latitude and longitude and converting it back in to a named place. This is Lucy Fellows' getting you to "look at the world her way".

Geotagging is the act of assigning geographic metadata to a photograph. As location information is increasingly stored in databases as a first-class data type, the way dates are, the phrase "geotagging" might now be considered a misleading. The term evolved at a time when few systems allowed users to index geographic data explicitly and so, taking matters into their own hands, they simply used existing tagging infrastructures to store, query and retrieve location information using ad hoc techniques. And the name stuck.

When users geotag photos on Flickr they, typically, first geocode an address to locate the place they want to say a photo was taken ,and Flickr then revers- geocodes that point in order to display the name of the place the photo was taken in. Geocoding is the art of inferring meaning from a multiplicity of written forms, but it is costly to perform and not a particularly efficient way to store and retrieve documents, photos in Flickr’s case, that have been geo-referenced, particularly when the initial geocoding may have been incorrect or used only to begin a more fine-grained positioning.

Which means: This is a story about naming things.

It is easy to get caught up in the rhetoric of the digital revolution and believe in what I think of as the illusion of addressability. Everything is just ones and zeroes and can therefore can be given a unique identifier, goes the argument. The IPv6 standard, for example, supports 2x128 possible IP addresses, leading many to claim every object created and every human born will be issued their own identifier and, in turn, the world will finally be connected in a seamless Web devoid of ambiguity. Unfortunately, the only "people" who relate to the world this way are robots.

Once upon a time, I worked at a small Internet service provider where we all shared the responsibility of fielding technical support calls. One day I answered the phone and was asked to explain to a person only just beginning to use the Web why the "" Web site was full of porn. This was a reasonable question, since most people were unaware that domain names are simply conversational short-cuts for the numeric addresses that actually identify sites on the Internet. I also explained that same "name" may exist in multiple domain spaces, none of which needed to know anything about the other, and that the actual United States government Web site was located at whitehouse .gov.

As it happens, the whitehouse .com Web site is no longer a porn site;  instead it is a news and information portal, but that only serves to illustrate the still-fugitive nature of "named places", even on the Web. That transience is further reflected in the commercial nature of the Internet which requires that ownership of a domain name be renewed at a yearly cost, and failure to do so is all it takes for another interested party to claim “your” name.

The problem is not that both whitehouse .com and whitehouse .gov have different IP addresses. The problem is that all names are shortcuts for interlocking and constantly evolving sets of ideas, assumptions and relationships that computer science struggles, and generally fails, to keep pace with. The same issues manifest themselves in daily life as the association of French Champagne producers, straddling both sides of the debate over globalization and the politics of identity and terroir, take out full page advertisements in U.S. magazines (New Yorker 20090119 pg. 7) decrying the use of the name "Champagne" by wines produced in California.

Long before the Internet exploded into everyone's lives we were collecting photos in shoeboxes and writing dates and place names on the back of each image; fussing over the time it took to do but eventually regretting the decision not to. Place is history, and names are a reflection of the experiences we share with close relations, the larger community and even our own past. For all its ambiguity and shifting meaning, the name we give a place is the air that a representation of that place breathes.

Geocoding a name, when someone searches for, or geotags, a photo, is only one-half of the problem. It allows you to fix a photo to a map, but how then do you connect that spot to memory and the history of the event?

How Does It Work?

We need to understand the difference between location and place. Computers and mobiles are very good at location, but we describe where we are as place, where culture meets location.  (Matt Jones)

Flickr works with a large database of places, historical and contemporary, called GeoPlanet. Every place is assigned a unique identifier called a Where On Earth (WOE) ID and contains information about its ancestors, children, sibling and neighbouring places. For scaling and performance reasons we opted to store only pointers to individual WOE IDs in the Flickr database itself. This meant that we had to boil the ocean a little in order to trim the number of possible members in any given location hierarchy. Eventually we settled on the following:

  • Neighbourhoods
  • Localities
  • Counties (optional)
  • Regions
  • Countries
  • Continents

It is worth noting that this is only one possible hierarchy, and some of the choices we've made were due solely to the mechanics of operating a site as large as Flickr. If only as an exercise, a critic might argue that our model is biased towards a philosophy of liberal economic governance and traditional capitalist land ownership. Although it would be wrong to ascribe that much motive to our actions (we simply started the geo project with an existing data source that had been originally developed for use by government agencies and worked with what we had), it is interesting to consider the possible facets, still present, in an otherwise seemingly rigid hierarchy.

A simple example is to contrast the way that Flickr and FireEagle (a Web application for collecting and sharing personal locative information) handle "localities" since the two sites share an almost exact hierarchy of places. Flickr treats anything with neighbourhoods as a locality, so in our model Duncans Mills, CA (pop. 84) and Mexico City (pop. 19M) are assumed to be the same "type" of place. FireEagle does not. If you authorize a third-party application to access information about your whereabouts at a city level, there is an expectation, assuming that you share an expectation that cities are "big", that your actual location will be suitably obfuscated (or "fuzzed"), and in a town of 84 people there's not a lot of room to get fuzzy in.

With all that in mind, when a Flickr user drops a photo on the map:

  • We calculate a search radius based on the map's zoom level, which is mapped to its corresponding place type in the Flickr hierarchy.
  • We query GeoPlanet for all the places of that type (say, neighbourhoods) that intersect the query radius.
  • We filter the list to only those places whose bounding box overlaps the center point, factoring in a degree of allowable fuzziness. If the center point falls outside a bounding box by a hundred meters, then it's usually still worth considering.
  • We then iterate over the second list measuring both the distance of the center point to the logical and political center point of each bounding box.
  • As a rule, the bounding box with the shortest distance wins, but the process is often messier and involves testing whether one bounding box is contained by another or whether a particular administrative relationship (the town where mail is typically delivered in the case of very small places) against the geographic reality.
  • If none of those tests succeed in finding a suitable location, we determine the parent of the place type we've just tested and try again until we’ve exhausted all the possible place types in our hierarchy. Once we have established a "root" location we then query for its ancestors and store each along with its WOE ID in the Flickr database.

For example, if you've tried to geotag a photo in the woods of Siberia and we can't find a matching town, we should at least be able to tell you the photo was taken in Siberia or, failing that, Russia. Sometimes, though, after all of that work we still choose to display the wrong place.

Synthetic Telepathy

Location is an 80/20 problem where the 20 really matters. (Aaron Straup Cope)

Until recently we've only ever been able to work with bounding boxes, a limiting function of the available data from our provider. Despite that, we have been surprisingly successful at mapping location to place even in the case of neighbourhoods. But there have always been mistakes, and no one is very tolerant of mistakes about "place". Never mind so-called disputed places (Kashmir, the West Bank, Cyprus, etc.): all neighbourhoods are "disputed" around the edges. This is often true of localities, as well. Our experience, reverse geocoding photos at Flickr, has been that there are few better ways to pick a fight than to tell someone what neighbourhood they are in and being wrong.

The problem of course is that even if we mapped every combination of latitudes and longitudes, multiplied by an infinite number of decimal points, to a single place, people still wouldn't agree on the answer. In the same way that a point is really just a very small bounding box, a single point is also a flattening of the history of that place and, ultimately, there is only so much human subtlety you can, literally, codify in to a computer program.

Instead of simply trying to keep pace with all of human history and prejudice as a series of cascading if/else statements, what if we laid our cards (the named places that we think a pair of latitude and longitude coordinates might be) on the table, and when we are wrong, give people the chance to tell us what they meant and to learn from that? What if the next time you geotagged a photo, we compared where we think that place is against the places that you've told us are nearby? If not you, then your contacts? What if every single person on Flickr points out that a neighbourhood, or town, is just plain wrong?

By adding a relatively small change to the site, allowing people to indicate that the place we had associated with a location was incorrect and allowing them to choose from a list of available options, we were able to better reflect their understanding of the world and begin to map facts on the ground rather than from on high. In the first week alone, we received one hundred thousand corrections! Depending on your point of view, this is either a testament to community-driven data, and so-called neo-geography, or proof that everything we've done to date was broken and wrong.

The Shape of Observation

"We should be mapping information that in some ways has been historically unmappable because it is 1) not valued or is 2) actively seen as threatening or is 3) simply too hard to map using traditional tools." (Anselm Hook)

I tell these stories because as of this writing Flickr has over 100 million geotagged photos, each of which has up to six unique place (WOE) IDs associated with it. Over time we've wondered: if we plotted all the geotagged photos associated with a particular WOE ID, would we have enough data to generate a mainly accurate contour of that place? Not a perfect representation, perhaps, but something more fine-grained than a bounding box. It turns out we can, effectively rendering the contour of all the points associated with a place into a recognizable shape, using software we developed called Clustr.

Clustr is a thin wrapper around the open source Computational Geometry Algorithms Library (CGAL) and uses a technique called "alpha shapes" to calculate the shape formed by a set of points:

Imagine a huge mass of ice-cream making up the space ... and containing the points as "hard" chocolate pieces. Using one of those sphere-formed ice-cream spoons we carve out all parts of the ice-cream block we can reach without bumping into chocolate pieces, thereby even carving out holes in the inside (eg. parts not reachable by simply moving the spoon from the outside). We will eventually end up with a (not necessarily convex) object bounded by caps, arcs and points. If we now straighten all "round" faces to triangles and line segments, we have an intuitive description of what is called the alpha shape.... (Tran Kai Frank Da, Mariette Yvinec)

 The results have been stunning, and while we can draw a near perfect outline of the United States or France or Texas using no other geographic information than the locations associated with photos, many, if not most, of the shapes we create look a little weird. Possibly even “wrong”. This is both okay and to be expected for a few reasons:

  • Sometimes we just don’t have enough geotagged photos in a spot to make it is possible to create a shape. Even if we do have enough points to create a shape there aren’t  enough to create a  shape that you’d recognize as the place where you live. We chose  to publish those shapes anyway because it shows both what we know and don’t know about a place, and it encourages users to help us fix mistakes.
  • We did a bad job reverse-geocoding photos for a particular spot and they’ve ended up associated with the wrong place. We’ve learned quite a lot about how to do a better job of it in the two and a half years we’ve been doing this, but human awareness is fickle and does not always lend itself to being formalized.
  • Sometimes, the data we have for trying to work out what’s going on is just bad or out of date, and we rely on users pointing out what is obvious to them but novel and unexpected to us.
  • We are not very sophisticated yet in how we assign the size of the alpha variable when we generate shapes. As far as we can tell, no one else has done this sort of thing so as with reverse-geocoding, we are learning as we go. For example, with the exception of continents and countries, we boil all other places down to a single contiguous shape. We do this by slowly cranking up the size of the ice cream scoop; this in turn can lead to a loss of fidelity. There is a lot left to learn.

Does the "shape" of Florida, or of Italy, include the waters that lie between the mainland and the surrounding islands? It’s not usually the way we imagine the territory that a place occupies, but the warping of the coast of Massachusetts by people taking, and geotagging, photos while on whale-watching boats is not an entirely inaccurate depiction of place either. On the other hand, including the ocean between California and Hawaii as "part of" the United States would be kind of dumb.

"Donut Holes"

They do not detail locations in space but histories of movement that constitute place. (Rob Kitchin, Chris Perkins)

More recently, while generating visualizations of these place shapes, we've noticed some interesting patterns. If we draw the shape of the city of Paris and then, on top of that, draw the shapes of all the city's child neighbourhoods, we see a richer and subtler definition of its boundaries.

The first outline maps roughly to the extremities of the RER, the commuter train that services Paris and the surrounding suburbs. This is a fairly accurate representation of the “greater metropolitan” area of Paris, reflected in both popular folklore and government administrivia as more and more people shift from rural to urban living. The rest, taken as a whole, follows closer to the shape of the old city gates that most people think of when asked to imagine Paris. Which one is right? Both, obviously!

Cities long ago stopped being defined by the walls that surround(ed) them. There is probably no better place in the world to see this than Barcelona which first burst out of its Old City with the construction of the Eixample at the end of the 19th century, and then again, after the wars of the 20th century, pushed further out towards the hills and rivers that surround it.

There are lots of reasons to criticize urban sprawl as a phenomenon, but sprawl, too, is still made of people who over time inherit, share and shape the history and geography they live in. Whether it’s Paris, Los Angeles, William Gibson’s dystopic “Boston-Atlanta Metropolitan Axis” (BAMA) or the San Francisco “Bay Area,” they all encompass wildly different communities whose inhabitants, in spite of the grievances harboured towards one another, often feel as much of a connection to the larger whole as they do to whatever neighbourhood, suburb or village they spend their days and nights in.

That’s one reason it’s so interesting to look at the shape of cities and see how they spill out beyond the boundaries of traditional maps and travel guides. In the example above, the shape for Paris completely engulfs the commune of Orly, 20 kilometers to the South of central Paris: this  makes a certain amount of sense. It also contains Orly airport which isn’t that notable except that Flickr treats airports as though they were cities in their own right; the realities of contemporary travel mean that airports have evolved from being simple gateways to capital-P places with their own culture, norms and gravity. So, now you have cities contained within cities which most people would tell you are just neighbourhoods.

As of this writing, we’ve finished rendering the third batch of shapes for the corpus of places in the Flickr database and looking ahead are wondering whether we should also be rendering shapes based on the “relationship” of one place to another. Rendering the shape of the child places for a city or a country would allow you to see a city’s “center” but also provide a way to filter out parts of a shape with low Earthiness (aka water) quotient, typically countries.

The issue is not to prevent, or correct, shapes that provide a false view, because I don’t think they do. Schuyler Erle, developer of the Clustr application, observed while we were getting all this stuff to work in the first place and testing the neighbourhoods that border the San Francisco Bay that they are really “the shapes of people looking at the city”. They are each different, but the same.

But maybe we should also map the neighbourhoods that aren’t considered the immediate children of a city but which overlap its boundaries. What if you could call an API method to return the list or the shape of a place’s “cousins”? What could that tell us about a place?

Communities of Authority

The “long here” that Flickr represents back to me is becoming only more fascinating and precious as geolocation starts to help me understand how I identify and relate to place. The fact that Flickr’s mapping is now starting to relate location to me the best it can in human place terms is fascinating ... but where it falls down it falls down gracefully, inviting corrections and perhaps starting conversation  (Matt Jones)

We could have released these shapes before the corrections project, but then it would have been little more than a closed cycle, where our misinterpretations of place were relayed back to our data provider and so on. By giving users the ability to signal their interpretation of place, we not only break the feedback loop, but also provide a way for those corrections to be fed back in to Flickr's reverse-geocoding engine to better geotag photos in the future: we use the wisdom of the community to give shape and nuance, and voice, to the authority of the dataset that we are working from.

As with any visualization of aggregate data, there are likely to be areas of contention. One of the reasons we’re excited to make the data, via the Flickr Application Programming Interface (API), available is that much of it simply isn’t available anywhere under a non-commercial license, and the users and the developer community who make up Flickr have a gift for building magic on top of the API so we’re doubly-excited to see what people do with it.

Clustr, the software used to generate shapes, was released under an open source software license and is designed to work with any set of latitudinal and longitudinal derived points. In the future we hope to add a feature to assign an abstract weighting to any individual point to affect how it is interpreted by the application. For Flickr, this weighting might be whether or not its associated photo was corrected or whether the location was offered as a suggestion by another user. Another limiting agent might be whether a photo was geotagged by a user who could be considered a resident of that place, rather than a tourist or visitor.

But more than that, we hope other projects will start to map the shape of their projects and share them with the wider community.

History Boxes

The (Minnesota Historical) Society has the largest collection in the universe of Minnesota fiction and many of these books create thinly veiled places based on the author's experience with an authentic local place. (Patrick Coleman)

What are the shapes of user-defined places? The first place you kissed your spouse? Napoleon's march in, and then back out, of Russia? Your daily commute? Does the shape of New York City's "ground zero" extend beyond the city blocks excavated after the World Trade Center towers fell, to the places that people ran to, or to the vantage point from which a person saw events unfold?

The Massachusetts Institute of Technology's SENSEable Cities project has been researching and visualizing the movement of tourists in Barcelona through the photos they've posted to Flickr, since 2007. What would it mean, not simply to plot those photos as a cloud of isolated events, but to give them shape, and meaning, as entirely new neighbourhoods or temporary cities in time, like Black Rock City which seems to emerge fully-formed out of the Nevada desert for the annual Burning Man event, only to disappear and "leave no trace" (except, as it happens for a lot of geotagged photos) ten days later?

Stamen Design's Oakland Crimespotting is an interactive map for visualizing and understanding crimes in the city of Oakland. By filtering incident reports by date and type, a viewer is able to see both the shape of criminal behaviour in the city and also the shape the city's response to it; for example, the seemingly clockwork intervals between no activity in a neighbourhood followed by nearly block-to-block reports of prostitution arrests. What is the shape of the history unseen in a place?

"Over 9, 000 Londoners lost their lives to V2 rocket strikes in World War 2," writes Tom Taylor, creator of the Web site. "Below are the five ... rocket strike locations" nearest to Westminster. Financial institutions and lenders may want to pattern the world with spending habits and agency, but I'd like to use the same tools to see the pattern of nearby. We all have friends who've sat in the same seat on the same airplane flying back and forth between destinations, only to have a third friend, years later with the aid of the Internet and a GPS-enabled device, see a photo of that seat and realize they are sitting only meters away. Not on the same plane, but in the same physical space that both planes occupied and the same place, the same anti-space, that anyone seated on a plane waiting to taxi from the terminal to the runway inhabits.

You are here, so say all the maps.

The Hammock of Interpretation

Place is history, and if the Internet is even half the "architecture of participation" that its supporter claim, then maybe history need no longer be written by the victors alone. Given the chance, what are the dinner-time, war-time and drunken kitchen-party stories that the places we have known would tell?

What would they name?


For Quotations

(Raisz, Fellows)




(Da, Yvinec)

(Kitchin, Perkins)




Cite as:

Cope, A.S., The Interpretation of Bias (and the Bias of Interpretation) . In J. Trant and D. Bearman (eds). Museums and the Web 2009: Proceedings. Toronto: Archives & Museum Informatics. Published March 31, 2009. Consulted