The Great(er) Bear - using Wikidata to generate better artwork

One of my favourite works of art is The Great Bear by Simon Patterson.

At first glance, it appears to be a normal London Tube map. But look closer...

Cool! But there is something about it which has always bothered me. Each Tube line represents a theme - therefore, a station at the intersection of multiple lines should be represented by someone who matches all of those themes.

For example, here's Baron's Court - the intersection of the Explorer line and the Saint line - represented by Saint Ursula.

She is just an saint - she has nothing to do with exploring. This artwork is wrong!

So, can we write something to query Wikidata to generate a more accurate artwork?
Because accuracy is my aesthetic.

Jump straight to the finished product and skip all the geeky details!

A brief guide to SPARQL

Wikipedia holds structured data about people and things. It uses SPARQL to query that data. It is a bit complex to use, but a valuable skill.

For example, this query finds people who are explorers and also saints.

SELECT DISTINCT ?person ?personLabel WHERE {
  ?person wdt:P106 wd:Q11900058 . # People whose occupation (P106) is explorer (Q11...)
  ?person wdt:P411 wd:Q43115      # People whose canonization status (P411) is sainthood (Q43...)

  SERVICE wikibase:label {
    bd:serviceParam wikibase:language "[AUTO_LANGUAGE],en" .
  }
}

The result is just four people. And that's where the problem starts. Simon Patterson chose categories for the lines which often don't have any intersections. There is no one who is an Italian Artist who is also a Saint and also a Footballer.

So, to create a more accurate artwork, we'll need different categories.

Nodes

One of the first things we need to do is understand the Tube map as a graph - with stations as nodes and lines as edges.

We want to know:

1. Which stations are on which lines
2. Which stations are on multiple lines
3. Which station has the most lines
4. How many stations are on each line

Thankfully Mark Dunne has done lots the hard work for us, and provided a great tutorial. Sadly, the data are about 5 years out of date.

Alternatively, the TfL API has lots of the information we need. Here's the call for all the stations on the Bakerloo line - https://api.tfl.gov.uk/line/bakerloo/stoppoints

Let's throw some Python down to grab the data we need. First, how many stations are there on the Bakerloo line?

import requests
r = requests.get("https://api.tfl.gov.uk/line/bakerloo/stoppoints")
stations = r.json()
count = len(stations)
print("There are " + str(count) + " stations on the Bakerloo Line")

Next, let's get the lines for each station:

import requests
r = requests.get("https://api.tfl.gov.uk/line/bakerloo/stoppoints")
stations = r.json()
 
for station in stations:
    stationName = station["commonName"]
    lineGroups   = station["lineModeGroups"]
    for lineGroup in lineGroups:
        modeName = lineGroup["modeName"]
        if (modeName=="tube"):
            lineCount = len(lineGroup["lineIdentifier"])
 
    print(stationName + "," + str(lineCount))

The line names can be found at https://api.tfl.gov.uk/Line/Mode/tube

Brief survey of the problem...

270 Tube Stations(!) across 11 lines. King's Cross St Pancras has the most lines - 6.

There a few anomalies in the data. It lists Edgware Road as two separate stations - even though it's really one station.

The same problem is present on Hammersmith and Paddington. Cleaning data is "fun"...

The categories are also challenging. This is how many times the Bakerloo line intersects with the other lines

    'bakerloo': {
        'circle': 3,
        'hammersmith-city': 1,
        'jubilee': 2,
        'metropolitan': 1,
        'northern': 4,
        'district': 2,
        'central': 1,
        'victoria': 1,
        'piccadilly': 1,
        'waterloo-city': 1
    },

That is - the Bakerloo line touches every other line at least once. As do the Northern, Central, and Jubilee lines. Those lines will need to contain some very broad categories.

Back to Wikidata

So, we want to replace each station's name with a human's name. We need attributes which are wide-spread enough to get good coverage in the data - and quirky enough to be interesting. I'd also like to keep some of the original categories:

I suspect there's a way to interrogate SPARQL to find a list of categories based on a graph - but I'm not clever enough to do that. I started off with an entirely arbitrary set of attributes:

• Academy Award Winners
• Left-handed people
• Nobel Prize Winners
• People born in London
• Educated at UEA (the university where my wife and I first met)
• Female Computer Scientists
• Saints
• Explorers
• Journalists
• Sinologues
• Comedians

Here's the query for Comedians who were educated at UEA and were born in London:

SELECT DISTINCT ?person ?personLabel WHERE {
  ?person wdt:P69 wd:Q1045828 .
  ?person wdt:P106 wd:Q245068 .
  ?person wdt:P19 wd:Q84
  SERVICE wikibase:label {
    bd:serviceParam wikibase:language "[AUTO_LANGUAGE],en" .
  }
}

One result - Doc Brown. There are no saints who have won an Oscar, and data about left-handed people is suspiciously absent. The categories will have to be completely rejigged.

Keep It Simple, Stupid

I figured the easiest thing to do would be to start from a well data'd individual and work backwards from there.

SELECT DISTINCT ?person ?personLabel WHERE {  
  ?person wdt:P108 wd:Q35794  . #Employed by Cambridge University #Bakerloo
  ?person wdt:P19  wd:Q84     . #Born in London                   #Circle
  ?person wdt:P463 wd:Q123885 . #Member of the Royal Society      #Hammersmith&City
  ?person wdt:P106 wd:Q121594 . #Professor                        #Waterloo&City
  ?person wdt:P106 wd:Q205375 . #Inventor                         #Metropolitan
  ?person wdt:P106 wd:Q81096  . #Engineer                         #District
  ?person wdt:P106 wd:Q4964182. #Philosopher                      #Piccadilly
  ?person wdt:P106 wd:Q11063  . #Astronomer                       #Victoria
  ?person wdt:P106 wd:Q170790 . #Mathematician                    #Jubilee
  ?person wdt:P106 wd:Q82594  . #Computer Science                 #Northern
  ?person wdt:P106 wd:Q188094 . #Economist                        #Central

  SERVICE wikibase:label {
    bd:serviceParam wikibase:language "[AUTO_LANGUAGE],en" .
  }
}

The result of that query is the inventor of steampunk, Charles Babbage!

We can do a reverse query. Given these people, which common properties do they have?

SELECT ?property ?propnameLabel ?value_Label
where
{ 
  wd:Q46633 ?property ?value . #Babbage
  wd:Q7259  ?property ?value . #Lovelace
  ?propname wikibase:directClaim ?property . # constrain to directClaims
  SERVICE wikibase:label { bd:serviceParam wikibase:language "[AUTO_LANGUAGE],en".
                         ?value rdfs:label ?value_Label .
                         ?propname rdfs:label ?propnameLabel .}
} order by ?property

(Thanks to TagishSimon for the help)

This is where things got trickier! Most of the major intersections didn't have any candidates other than Babbage - truly a Renaissance Man! - so I expanded "born in London" to "born in the UK".

We can't use ?person wdt:P27 wd:Q145 becuase that only covers the current United Kingdom - not The Kingdom of Great Britain (1707–1801) nor The United Kingdom of Great Britain and Ireland (1801 to 1927)

The correct query seems to be ensuring the place of birth is within the current administrative territory of UK:

?person wdt:P19 ?pob . ?pob wdt:P131* wd:Q145 .

But you can also use a UNION

{?person wdt:P27 wd:Q145} UNION {?person wdt:P27 wd:Q174193} UNION {?person wdt:P27 wd:Q161885} .

That got closer - but still not enough.

Unions

Wikidata is fickle. Someone may have an occupation as a "computer scientist" or they may work in the field of "computer science". SPARQL eschews the or operator, and uses UNION:

{?person wdt:P101 wd:Q21198} #Field of Work CS
UNION
{?person wdt:P106 wd:Q82594} #Occupation CS

I'm beginning to see why the original artist was more liberal in his accuracy!

Sorting

I want the map to contain notable people. There are a couple of ways to assess the "notability" of a Wikidata subject. I've chosen to use "sitelinks" - that shows how many languages their article is available in. It's a crude, but quick method.

Here's it in action:

SELECT DISTINCT ?person ?personLabel ?sitelinks WHERE {
  ?person wdt:P106 wd:Q205375 . #Inventor
  ?person wikibase:sitelinks ?sitelinks .

  SERVICE wikibase:label {
    bd:serviceParam wikibase:language "[AUTO_LANGUAGE],en" .
  }
} ORDER BY DESC (?sitelinks)

Correcting for Bias

Wikipedia has an acknowledged male bias. So I used SPARQL's FILTER property to great effect:

FILTER ( !EXISTS{ ?person wdt:P21 wd:Q6581097 }) 

It says to return anyone without the sex/gender of "Male". (Yes, I know things are a bit more complicated than that - but this is a good way to return women, intersex people, agender, transgender folk etc).

If no non-men were returned, I repeated the search but omitted the filter.

Because I used "Born in the UK" as a filter, there is probably a bias towards white people. And people who become Professors or members of the Royal Society may also be the product of a biased society. There are many other filters and categories I could have chosen - and I hope some of you will create maps for your own cultures and societies.

P-p-p-pickup Some Python

There are several Python libraries for SPARQL, I used sparqlwrapper.

Here's a sample query

from SPARQLWrapper import SPARQLWrapper, JSON
 
sparql = SPARQLWrapper("https://query.wikidata.org/sparql")
sparql.setQuery("""
    SELECT DISTINCT ?person ?personLabel ?sitelinks WHERE {
        ?person wdt:P106 wd:Q81096 . #Engineer
        ?person wdt:P20  wd:Q84    . #Died in London
        ?person wikibase:sitelinks ?sitelinks .
        SERVICE wikibase:label {
            bd:serviceParam wikibase:language "[AUTO_LANGUAGE],en" .
        }
    } ORDER BY DESC (?sitelinks)
""")
sparql.setReturnFormat(JSON)
results = sparql.query().convert()
 
for result in results["results"]["bindings"]:
    print('%s %s %s' % (result["person"]["value"], result["personLabel"]["value"], result["sitelinks"]["value"]))

Plotting onto an image

Let's leave who and what we select for now, and work out how we draw the eventual results.

We could do all sorts of clever things plotting out locations - but I decided to cheat!

There is a fully semantic SVG of the tube lines (Thanks to Oliver O’Brien for finding it).

I removed all the lines I didn't want, then I was able to search & replace station names with my preferred text.

Search And Replace

Well... sort of! Charing Cross Station has an ID of 940GZZLUCHX.

On the map it is:

<g id="s-940gzzluchx_label">
    <g id="s-940gzzluchx_label_1_">
        <text transform="matrix(1 0 0 1 515.8999 487.6963)">
            <tspan x="0" y="0">Charing</tspan>
            <tspan x="4.3" y="4.5">Cross</tspan>
        </text>
    </g>
    <polygon id="s-940gzzluchx_nr" fill="#EE3124" points="514.3,487.7 512.6,486.9 515,486.9 515,486.4 512.5,486.4 513.8,485.8 515,485.8 515,485.3 513.8,485.3 512.2,484.6 511,484.6 512.7,485.3 510.2,485.3 510.2,485.8 512.7,485.8 511.4,486.4 510.2,486.4 510.2,486.9 511.5,486.9 513.1,487.7 "/>
</g>

Whereas the SVG element for North Wembley is just:

<text id="s-940gzzlunwy_label_2_" transform="matrix(1 0 0 1 282.1489 289.1079)">North Wembley</text>

Again, I'm not quite clever enough to work out a way to reliably find the inner text for an element which may be inside (or not) several other similarly named elements.

So a lot of repetitive search-and-replace it is. *sigh*

I also need to manually place some of the station names, because they're a different length to the originals. *double-sigh*

Putting it all together

I have great pleasure in revealing to you "The Great(er) Bear"!

See the full sized version.

Copyright

OK gang, turns out that copyright law is even trickier than computer code! I've spoken to Simon Patterson and he is happy for me to host a not-for-profit version of this piece of art which is heavily indebted to his original.

TfL has been litigious in the past when it comes to derivative maps. I tried contacting them several times, but didn't receive any clear answers as to whether I could do this.

The data that I used to generate the art is "Powered by TfL Open Data" and provided under OGLv2. It may contain OS data © Crown copyright and database rights 2016.

The original font is tightly controlled. So I've used a freely available font called Hammersmith One which is broadly similar.

Lots of people create modified tube maps:

• The Literary Tube Map - putting authors on the map
• Tube Map of Beer Prices
• Decibel levels on the Tube
• House Price Tube Map
• And no doubt many more.

If you want to build your own version of my modified map, all the data are on my GitLab!

Details

Here are a few interesting close-ups of the map - they may be different from the final version.

Errata

• The data in Wikidata may be incorrect or incomplete.
• I originally didn't restrict it to just humans! So a few weird entries snuck in. Using ?person wdt:P31 wd:Q5 . corrected that. But I'm wondering if anyone on the map is fictional...!
• Due to timeouts and my crappy coding, I ran the code over several passes on different days. If you run the code, you might get different results.
• I didn't use people's names in their original language, I had to back-fill them. I probably missed some. I should have used P1559.
• Even after lots of jiggling of categories, one or two stations kept coming up blank. So I manually added in a few people. Can you spot who they are?
• Some people's names were too long for the allotted space, so I have swapped a few people around. Better code would try to keep name length as close to the original as possible.
• There's no (intentional) ordering. It might be nice to put people on the line in order of, say, year of birth.
• Similarly, there's almost no relation between the people and the places. Although I've contrived to put the author of Mary Poppins somewhere special!
• The Hammersmith One font only has a basic set of characters - so non-European languages (and some accents) are in the default font.
• The Elizabeth Line / CrossRail hasn't opened yet. I suspect it will be much harder to produce a new map once it goes live. Similarly the DLR and Overground lines are excluded.
• The SVG renders well in Firefox, and seems to work OK in Chrome. Please let me know of any glitches.
• I've also added a couple of Easter Eggs. Enjoy finding them!

Thanks

Mainly to my wife, Liz, for being very patient with me while I swore at my code.

I am indebted to Simon Patterson for his incredible and inspirational artwork. When it was created in 1992, Wikipedia did not exist. Linked Data stores were in their infancy. It would have been close to impossible to create a semantically correct map. Nothing in my version is intended to take away from Patterson's work and creativity.