Graphs and visualization

A lot of the joy of digital humanities comes in handling our material in new ways, so that we see things we wouldn't have seen before. Quite literally.

Some of the most useful tools for DH work are graphing tools! Today we will look at the basics of what a graph is and how you might build one, both manually and programmatically, and then name the tools to look into if you want to know more.

Tools you'll need

• Graphviz: http://www.graphviz.org
• The graphviz and sphinx modules for Python: pip install graphviz sphinx

So what can you do with graphs?

You can visualize relationships, networks, you name it.

http://ckcc.huygens.knaw.nl/epistolarium/#

The DOT graph language

It's pretty easy to start building a graph, if you have the tools and a plain text editor. First you have to decide whether you want a directed or an undirected graph. If all the relationships you want to chart are symmetric and two-way (e.g. "these words appear together" or "these people corresponded", then it can be undirected. But if there is any asymmetry (e.g. in social networks - just because Tom is friends with Jane doesn't mean that Jane is friends with Tom!) then you want a directed graph.

If you want to make a directed graph, it looks like this:

digraph "My graph" {
    [... graph data goes here ...]
}

and if you want to make an undirected graph, it looks like this.

graph "My graph" {
    [... graph data goes here ...]
}

Let's say we want to make that little two-person social network. In graph terms, you have nodes and edges. The edges are the relationships, and the nodes are the things (people, places, dogs, cats, whatever) that are related. The easiest way to express that is like this:

digraph "My graph" {
    Tom -> Jane
}

which says "The node Tom is connected to the node Jane, in that direction." We plug that into Graphviz, and what do we get? Let's use a little iPython magic to find out.

We are going to use an extension called 'hierarchymagic', which gives us the special %%dot command. You can get the extension by running this commands in a terminal window:

ipython -c '%install_ext http://students.digihum.ch/hierarchymagic.py'

Once that is done, here is how it works.

Then you do this every time you want to make a graph. The -f svg says that it should make an SVG image, which is what I recommend.

Now maybe Tom has a friend too:

digraph "My graph" {
    Tom -> Jane
    Ben -> Tom
    Tom -> Ben
}

...And so on. But what if we do want a nice symmetrical undirected graph? That is even simpler. Instead of digraph we say graph, and instead of describing the connections with -> we use -- instead. If we have a model like Facebook where friendship is always two-way, we can do this:

graph "My graph" {
    Tom -- Jane
    Ben -- Tom
}

Note that we don't need the third line (Tom -- Ben) because it is now the same as saying Ben -- Tom.

Since this is an undirected graph, we want it to be laid out a little differently (not just straight up-and-down.) For this we can specify a different program with this -- -K flag. The options are dot (the default), neato, twopi, circo, fdp, and sfdp; they all take different approaches and you are welcome to play around with each one.

Of course, this would hardly be fun if we couldn't do it programmatically!

Building graphs with Graphviz + Python

Now we are going to make a few graphs, not by writing out dot, but by making a graph object that holds our nodes and edges. We do this with the graphviz module.

We make a new directed graph with graphviz.Digraph(), and a new undirected graph with graphviz.Graph().

Let's make a social network graph of five friends, all of whom like each other. But instead of typing out all those

Anna -> Ben

sorts of lines, we will let the program do that for us.

And here is a little iPython magic function so that we can actually make the graph display right here in the notebook. This means that, instead of copy-pasting what you see above into a new cell, you can just ask IPython to do the copy-pasting for you!

Don't worry too much about understanding this (unless you want to!) but we will use it a little farther down. You can ignore the lines about "Couldn't evaluate or find in history" - that seems to be a little IPython bug!

Basic usage for the Graphviz python library

So here is a short summary of what we did above that you will want to remember:

Labels and IDs

When you are making a graph, it is important that every node be unique - if you have two people named Tom, then the graph program will have no idea which Tom is friends with Anna. So how do you handle having two people named Tom, without resorting to last names or AHV numbers or something like that?

You use attributes in the graph, and specifically the label attribute. It looks something like this:

graph G {
    Tom1 [ label="Tom" ]
    Tom2 [ label="Tom" ]
    Tom1 -- Anna
    Tom1 -- Tom2
}

Before this, we only named our nodes when we needed them to define a relationship (an edge). But if we need to give any extra information about a node, such as a label, then we have to list it first, on its own line, with the extra information between the square brackets.

There are a whole lot of options for things you might want to define! Most of them have to do with how the graph should look, and we will look at them in a minute. For now, this is what we get for this graph:

Notice, in this, that Anna still popped into existence when we referred to her in a relationship. But in the real world, we will probably want to declare our nodes with (for example) student numbers as the unique identifier, and names for display in the graph.

Styling the graph

We can also define how the graph looks, or how nodes look, or how edges look, through attributes!

digraph G {
    node [ shape="plaintext" fontcolor="red" ]   # Define the look of all nodes.
    mother [ label="Tara" ]
    father [ label="Mike" ]
    child  [ label="Sophie" ]
    work   [ shape="house" color="black" fillcolor="yellow" ]
    school [ shape="house" color="black" fillcolor="green" ]

    mother -> child [ label="is mother" ]
    father -> child [ label="is father" ]
    mother -> work  [ label="goes to" ]
    father -> work  [ label="goes to" ]
    child -> school [ label="goes to" ]        
}

And here's how we do that in python, and what we get...

All done! There are a huge number of styling attributes - ways to control line thickness, color, shape, graph direction, and so on. They are documented in eye-watering detail here:

http://www.graphviz.org/content/attrs

Have fun!