A Tutorial

By Clement Levallois, Erasmus University Rotterdam

V 1.1 - 2013
Bio notes

• Education in economics, management, history of science (Ph.D.)

• Since 2008, turned to digital methods for research.

– data visualization, network analysis, natural language processing, web
applications and more.
– Member of the “Gephi Community Support” team
– Gephi certified trainer
https://marketplace.gephi.org/service/data-analysis/

• Contact, feedback welcome: on twitter @seinecle or

www.clementlevallois.net

Clement Levallois
 Table of content
Presentation of Gephi 4 Detecting communities
Setup 8 & visualizing them 32
Adding plugins 9 Tips to clean up the appearance 34
A note on terminology 10 Gephi: what’s the output? 35
--- start case 1 ---- 12 Export your visualization! 36
Importing the dataset 14 --- end case 1 ---
Import report 15
Saving a Gephi project 17 The “hairball” problem 41
Mouse controls 18 Tackling the hairball problem 42
Rearranging windows, tabs, panels 19
Switching on labels 20 Helper on importing CSV files 46
What a layout is 21 Helper on filters 47
Example of a simple layout 22 Helper on ranking panel 48
An insightful layout: Force Atlas 2 24 Helper on centrality 49
How do force-based layouts work 26
Making nodes bigger 27 Cheat sheets 51
Visualizing attributes in your data 28
Partitioning or ranking? 29 References 56
Example of partitioning 30
Visualizing new attributes with Gephi 31

Clement Levallois
Gephi
• Created in 2008 by a core team of 4 French
computing engineers inspired by a professor.

Clement Levallois
Gephi
• In 2013,
– 11 core developers
– 15 developers of plugins
– 4 Google Summer of Code
– One Java Duke Award
– 210,000 downloads in the past year

• Localization is available in French, Spanish, Japanese, Brazilian

Portuguese, Russian, Chinese and Czech.

• About 1,000 academic papers citing Gephi

• Very active forum (www.forum.gephi.org)

Clement Levallois
Gephi
• A software written in Java for Mac, PC and Linux

• A headless version (the Gephi Toolkit) to generate

graphs automatically via API

• An ecosystem of plugins and related tools

• A commitment to open source, quality, and freedom to

use.

Clement Levallois
A note on the slides

Signals an important feature,

or an error to avoid.

Clement Levallois
Setup
• Check list:
- Make sure you have the latest version of Gephi installed

- In the menu, choose Help -> Check for updates

- You need a mouse with a scrolling wheel, to zoom ( )

- In the menu of Gephi, go to Tools -> Options -> Visualization -> OpenGL
There, increasing the anti-aliasing factor can improve the quality of the view
on your screen.

- We are going to download some stuff. Check your internet connection.

Clement Levallois
Adding plugins
• In the menu, go to
Tools -> Plugins -> Available plugins

• Select with the tick box the following plugins:

Circular Layout, OpenOrd Layout, GeoLayout, alphabeticalsorter

• Click install and follow the instructions.

Note: You will remark there is a “Plugins” item in the menu as well,
which is empty. That’s a secret feature. Or just a very misleading UI design.

Clement Levallois
A note on the terminology

X Y
Y was born in …
Y’s marital status is …
Y’s latitude is …
Y’s longitude is …
An “edge”
A “node” (or a tie, or a link…)
(also called a vertex, Nodes attributes
plural vertices). (NB: edges can have attributes too!)

A directed network An undirected network A weighted network

(the direction of the relation (the direction of the relation does (the edges have a “strength”
matters) not matter) represented by a numerical value)
Clement Levallois
 The plan for today
• 1st exercise: one guided tour of a dataset from
start to finish
Here, let’s keep interruptions to a minimum, the point is to
practice a typical workflow from A to Z

• 2nd exercise: you work in autonomy on one

dataset
Here we will spend time looking at more custom and advanced
features.

Clement Levallois
1st practice: workflow in Gephi
• Import the network

• Spatialize the network

• Visualize attributes of your data

• Visualize attributes created by Gephi

• Export

Clement Levallois
The dataset
• Coappeareance network from Les Miserables, by
Victor Hugo

• 2 characters are connected if they are mentioned close

to each other in the text.

• Dataset originally provided byThe Stanford GraphBase,

http://www.ctan.org/tex-archive/support/graphbase/

• Download the file from

www.clementlevallois.net/gephi/tuto/miserables.gexf

Clement Levallois
1. Importing the dataset in Gephi

Clement Levallois
 2. Import Report
This tab gives additional
This report provides information on the file you just opened info on the attributes
found in the graph –
useful to spot errors in
the file

If necessary, the
graph will be
resized to fit in the
window.
Should an edge A-B
be seen actually as
A->B? You decide it
here.
If you want
to merge the
graph with
Number of one that is
nodes found already open
in Gephi
Number of
edges found

Is the graph time If your have a time-

evolving? stamped graph.
Importing the same
Is the graph a meta- graph with different
graph? time-stamps creates
Often, networks files are simply a list of nodes (ex: A, B and C), and a list
*we will ignore a “time evolving”
of edges connecting these nodes (ex: A –B, B-C) . What if Gephi finds an
hierarchical graphs in (dynamic) graph
edge A-D in the file? Since D is not present in the list of nodes, you are
this tutorial* given here a choice to add it now. If not, the edge A-D will be ignored. Clement Levallois
3. The graph, opened

Clement Levallois
4. Saving a Gephi project
• Gephi can save projects under the .gephi extension.
– These files can only be opened in Gephi, and contain all
information about your project. Click File -> Save.
– There are a number of reports of this files to be unstable,
so always also export your network as indicated just
below.

• The bare network itself can be exported as a file, in

different formats.
– gexf and graphML are two network formats which can
store attributes for nodes and edges.
=> Click File -> Export -> Graph File to save your network.

Clement Levallois
5. Mouse controls
• Left click: accomplishes the action you previously
selected

• Right click: move the graph around (“panning”)

• Scrolling wheel: zooming in and out

• “Alt” key + keep left click down while on the

graph: moves the graph in the 3rd dimension
(gives an effect of depth and perspective)

Clement Levallois
6. Rearranging windows, tabs, panels

My experience: don’t use

this “Configure Window”
menu.

Just drag, drop and resize

the tabs with the mouse
as you find most
convenient.

Try to maximize the space

for the main tab – the
one where the graph is
displayed!

Clement Levallois
7. Switching on labels
• Please refer to the cheat sheet 1 in appendix of this slide
deck.

• Notice the panel at the bottom of the screen, and its “Labels”
tab.

• Edges and nodes can have labels but in most use cases, only
nodes have them.

Tick here to display node labels Clement Levallois

8. What a layout is
• A layout is simply a
procedure to position
nodes on the screen
according to some
logic.
• Many procedures
exist – you could
create your own if
you’d wish

Clement Levallois
9. Example of a simple layout
• Select the “Circular
Layout” in the drop
down menu of
layouts.

• Don’t change any

setting and click “run”

Clement Levallois
10. Example of a simple layout
(continued)

The “circular layout” applies the following procedure: “spread nodes on a circle, at
an equal distance from each other.”

Can we distribute nodes according to different procedures, yielding more insights?

Clement Levallois
11. A insightful layout: Force Atlas 2
• Select “Force Atlas 2”
in the drop down
menu of the layouts.

• Change the scaling

parameter to “200”
(not necessary, actually…)

• Run it!

Clement Levallois
12. Force Atlas – the result

The result is that densely connected nodes are grouped in the same regions.

This property is very useful to interpret a graph:

- groups of nodes can be spotted immediately,
- Relations between groups are also visible.

Recent works have demonstrated analytically this result (that with certain kinds of
layouts, densely connected groups of nodes find a visual translation in a 2D plane).
See Noack (2009) and Waltman, van Eck and Noyons (2010) in the reference list.
Clement Levallois
13. How do force-based layouts work
• “Nodes repulse each other like magnets, while edges
attract their nodes, like springs”

• “These forces create a movement that converges to a

balanced state.”

• Force Atlas 2 specificity:

“[We] tweak the repulsion [“magnet”] force so that poorly

connected nodes and very connected nodes repulse less”

(see Jacomy et al. 2012 in the reference list).

Clement Levallois
14. Making nodes bigger
• Please refer to the cheat sheet 2 in appendix
to this slide deck.

• Right-click on the icon marked “size” and set

the number to 15 or higher. Validate.

• Then left-click on the same “size” icon.

Clement Levallois
 15. Visualizing attributes in your data
• Nodes in our example network have one
attribute: the gender of the character in “Les
Miserables”.

• We can overlay this information on the network

by color-coding the nodes according to the values
of this attribute.

• Note: we have 2 values for this attribute (male or

female), but this could be many more.

Clement Levallois
16. Partitioning or ranking?
• Partitioning is this: Male
Fem
ale

Works for attributes that classify nodes in categories: “male or female”, “East,
West, South or North”, “country of residence”. Each category is represented by a
different color, as above.

• Ranking is this:
28 35

or this: 28 35

Works for attributes that are gradual, not categorical (age, etc… any numerical
attribute).
The graduation is represented visually either by bigger sizes, or by changes in
shades of color (from light to dark, or from light colors to warm colors, usually).

Clement Levallois
17. Example of partitioning
• In the “partition” tab, hit the refresh button
to make sure Gephi loads all avail. attributes.

• In the drop down menu, we then select

“Gender” which is anyway the only available
attribute available for partition in our case.

• Click on “apply”!

• Note that if we had had a categorical attribute

for an edge (relation btw 2 character: friend,
enemy or neutral?), we could have colorized
edges according to the same principle.

• Note that if we wanted to group all nodes from

the same category as one big node, we would
have gone here. (that’s hierarchical graphs)
Clement Levallois
18. Visualizing new attributes with
Gephi
• Gephi can compute metrics, such as “how central
is each node” in the network.

• The key insight is: these metrics are attributes to

nodes (or edges), and can in turn be visualized
just as any other attribute.

• Visualizing these metrics (instead of reading them

in a tabular form as is usually done) stimulates /
speeds up insight.

Clement Levallois
 19. Detecting communities
& visualizing them
• Go to the statistics panel and run the Modularity algorithm.

• This algo identifies relatively densely connected groups of nodes in the network.

• As a result, each node ends up being attributed to one of these groups (called
“class” here).

• The same way we have colorized nodes by their genders, we can colorize them by
the group (“class”) they belong to.

• Go to the partition tab, hit the refresh button, select “modularity class”, and
“apply”!

• Note: by clicking on the colored squared next to each category in the partition tab,
you can change and choose the colors you prefer. In a future release of Gephi, this
color choosing mechanism will be improved further.

Click on it to choose
a different color
Clement Levallois
20. Visualization of communities:
the result

Clement Levallois
21. Tips to clean up the appearance
• Apply the layout “Label adjust”
– Find it in the drop down menu of the layouts

• Switch to a black background?

– Check the bottom left corner of cheat sheet 2

• Decrease the thickness of edges?

– A slider makes it easy

Clement Levallois
22. Gephi: what’s the output?
• The next slide shows how to export your viz to a picture file. Is it
what Gephi is for, ultimately?

• In my view, the value added is before all in the insights you gain in
the process of exploring the viz interactively, and iteratively, in
Gephi.

• Exporting the view to a file or a webpage meets different needs,

such as communicating the insights you gained.

• Jer Thorp, a reference in data visualization, has a short blog post in

the Harvard Business Review making a closely related point:
http://blogs.hbr.org/cs/2013/04/visualization_as_process.html

Clement Levallois
23. Export your visualization!
• Directly from the overview panel with the screenshot option (see
cheat sheet 2 ).

• Or in the preview panel (see cheat sheet 5 ).

• Why do the views in the overview and preview panels don’t exactly
match?
– Visualizing a graph in real time as in the overview panel requires a
technology (OpenGL) which is incompatible with an export in the pdf
or svg file formats.
– The preview panel provides this bridge towards pdf and svg, by using a
different technology to render the graph.
– This difference in technology means that some visual features in the
overview panel don’t translate exactly to the preview mode.

Clement Levallois
24. Les Miserables network
just before export to pdf

Clement Levallois
 END CASE 1
• It would remain to cover:
– example of a ranking
– example of a filtering
– dynamic networks
– import formats
– plugins (huge chunk!)
– the Gephi toolkit
– …

• Yet, practice is essential too. The 2nd part is a semi-

guided exercise where some of the features listed
above will be evoked.

Clement Levallois
 CASE 2:
Digital humanities on Twitter
Twitter data collected with NodeXL

tutorials to collect twitter data:

http://mashe.hawksey.info/2011/09/twitter-network-analysis-and-visualisation-ii-nodexl/
http://social-dynamics.org/gathering-a-twitter-users-network-data-with-nodexl/
http://sunlightfoundation.com/blog/2012/05/24/tools-for-transparency-a-how-to-guide-for-
social-network-analysis-with-nodexl/

Clement Levallois
1. The original dataset
• Nodes: Twitter users who mentioned “digital
humanities” or #digitalhumanities in the on April
3, 2013 and the few days before.

• Edges: A connection between two twitter users if:

- One replied to each other
- One mentions another
- One follows another

• Total: 793 nodes and 17,215 edges

Clement Levallois
 2. The “hairball” problem

The dh twitter
network:

- spatialized with
Force Atlas 2

- modularity algo
applied

- partitioned
according to the
communities
found by the
modularity algo

One area of interest (the spike), but the rest is a big hairball. Not insightful!
Clement Levallois
3. Tackling the hairball problem
• The basic problem is that we have too many connections
=> If I mention somebody in just a single tweet, I’ll be connected
to this person on the graph.

• Can we weed these circumstantial relations out and retain

only those that are most meaningful?

• The approach of Gaze:

– Two nodes will be connected not if they have a direct relation,
but if they have enough relations in common with the other
nodes in the graph.
=> I’ll be connected to another person on the graph only if we are
strongly similar in terms of who we connect with.

Clement Levallois
4. The reworked dataset
• Original network is processed with Gaze
– Check www.clementlevallois.net/software.php
– Only the strongest connections are preserved.
Isolated nodes are deleted.
– Edge weight goes from 0.29 (small similarity) to 1
(absolute similarity).
• End up with 467 nodes and 14, 645 edges
• Find the original and reworked datasets at:
www.clementlevallois.net/gephi/tuto/dh.zip

Clement Levallois
 2. Task:
• Import
• Spatialize
• Compute modularity
• Partition according to communities
• Rank nodes by number of followers
• Compute centrality
• Rank nodes by their centrality

• What do you find?

⇒Export to a picture file.

Clement Levallois
• You work in autonomy

• I go from computer to computer to help out

• The following slides provide additional tips.

Clement Levallois
a) Helper on importing CSV files
• See cheat sheet 4 for reference
• Close all open projects, create a new empty one
• Go to the data laboratory
• Click on “import spreadsheet”
• Upload the files
– dh [Nodes].csv
– dh [Edges].csv
• Choose careful the type of the attributes! (String,
Integer or Float)
Clement Levallois
b) Helper on filters

A screen shot to help on the filter procedure:

We combine 2 filters:
- One which filters out (hides) edges that have a
weight under a certain threshold

- One that filters out (hides) all nodes not

belonging to the biggest group of connected
node (“giant component”). This filter is simply
used to avoid the visual clutter of having isolated
nodes float around the giant component.

- To combine the filters, nest one inside the other

by dragging and dropping them from the list of
filters above.

Clement Levallois
 c) Helper on ranking panel

The “spline” function is

useful to rank an attribute in a
non-linear way.

If all tweeter users have less

than 10,000 followers and
one has 200,000, the ranking
procedure would risk to see
all users completely dwarfed:
all nodes very small, and one
giant node.

Using a spline, we can correct

for that by deciding to
increase the size of the node
more rapidly for lower values
of the attributes.

Mathematically the ranking

will be biased, but the human
eye will be now able to
discern more subtile
differences.

Clement Levallois
d) Helper on centrality
• Centrality measures are different ways to
capture the elusive notion of being “central” in a
network.
• Betweenness centrality is one such approach:
– Among all the shortest paths connecting any two
nodes in a network, the most central node is the one
which lays on most of these paths.
A B C

In this trivial example, B is central because it sits on the

shortest path from A to C. A and C score 0 because they
don’t lay on any shortest path.

More on centrality measures: http://en.wikipedia.org/wiki/Centrality

Clement Levallois
Appendix

Clement Levallois
Gephi Cheat Sheet 1
The Overview Panel
Where all the functions are available to explore the network visually.

Partition Statistics
Computes metrics on the network.
These metrics are recorded, and can then be
(tab hidden here)
(tab hidden here) used to be displayed on the graph. Ex: compute
For categorical the centrality of nodes. Then use the ranking
attributes. panel to make central nodes, bigger.
Example:
to color all nodes
representing males
in yellow, and all
nodes representing
females in green Filters
To hide or display

Ranking only part of the

network.
The “library” and
For numerical its folders contain
attributes. the filters. For
Example: the example, filter
older the out nodes which
individual, the have less than 3
bigger the node. edges. Or filter
Works with out edges which
gradients of have a weight
colors too (ex: above some
the older the value.
indiv., the pinker
it is). Drag and drop
the filter you
choose in the
“queries”
window. Several
Layout filters can be
combined (ex:
Changes how the filter out male
network is spread indiv. that have
on the screen. less than 3
Select one of the connections to
layouts in the others).
drop-down menu
and apply it to
General settings for the appearance
see how the
position of nodes of nodes, edges and labels
is changed. This box of settings can be shown / hidden with the little arrow circled in red on the top right.
Here, you can set if nodes are visible in 2D or 3D, what is the default color of edges, etc.
The “labels” tab is particularly useful: should they be displayed or not, and at which size. Clement Levallois
Gephi Cheat Sheet 2 Focus on the icons of the Overview panel
Functions which are less frequently used have been grayed out

select individual nodes

select a region
move individual nodes
change the color of a node
change the size of a node
change the color of a node and its neighbors
add one new node
add one new edge
finds shortest path between 2 nodes
change the color of a node and its neighbors – edge weight aware.
display the information of the selected node

resize the graph to the size of the window.

changes the color of all nodes to a unique color. Right click to define the color
resets the size of all nodes. Right click to define this size.
colors all labels in a unique color
??
resets the size of all labels.

1 2 3
change edge
More label settings:
thickness
export a
1, the 3 buttons on the left:
How to memorize all
screenshot.
Click the
from left to right,
switch on or off the
- label size – should it track the node size?
- label color – should it track the node color?
these icons??
arrow for display of: - label font All these controls are also
resolution - nodes labels
settings - hulls (not implemented yet)
available with a more explicit
2, the slider: label size
switch - edges description in the panel here.
background - edge of the color of Once you know these controls well, the
the source node 3, the 2 buttons on the right:
color! (useful - label color icons are a quick way to access them.
- edge labels
for prints) - text to be displayed in the label Clement Levallois
Gephi Cheat Sheet 3
NODE VIEW
The Data Laboratory
Where the numerical and textual data for nodes and edges can be examined and modified.

Import function
Opens a dialog window to import nodes and
edges from a csv file into Gephi
To switch
between views
of nodes &
Extra
edges columns
Each node can
have extra
information,
besides its id
and label.
3 default This extra info is
columns written in
additional
for nodes columns.
Node: simply a
Example:
copy of the
here, each node
label column
is characterized
(or the id if
by a number,
there are no
recorded in a
labels).
column we
choose to call
Id: the unique
“Modularity
identifier of the
class”.
node
Columns can
Label: the name
contain
of the edge
numbers, text
which will be
or booleans
displayed next
(true / false).
to it if we
choose to.

Helper functions to quickly edit columns

Clement Levallois
Gephi Cheat Sheet 4
EDGE VIEW
The Data Laboratory
Where the numerical and textual for nodes and edges can be examined and modified.

Import function
Opens a dialog window to import nodes and
edges from a csv file into Gephi
To switch
between views
of
nodes & edges
Extra
6 default columns
columns Each edge can
have extra
for edges information,
Source and Target: besides its id and
the two connected label, type and
nodes forming the weight.
edge.
For example
Type: here, I added a
Is the direction of column to
the edge characterize the
meaningful? connection
between the 2
Id: the unique characters of the
identifier of the Miserables:
edge. friends or
enemies in the
Label: the name of novel?
the edge which will
be displayed next to
it if we choose to.

Weight: how
“strong” is the tie
between the two
nodes forming the
edge? This is a
numerical value.
Helper functions to quickly edit columns
Clement Levallois
Gephi Cheat Sheet 5
The Preview panel
Where you make final adjustments before exporting your visualization to an image file (PDF, SVG or PNG)

Load or save
parameters

1. Setting
the
parameters

3. Export to a 2. Hit refresh!

picture file format After changing a parameter, you must hit
“refresh” to see the effects. Clement Levallois
References
Bastian M., Heymann S., Jacomy M. (2009). Gephi: an open source software for exploring and
manipulating networks. International AAAI Conference on Weblogs and Social Media.
http://gephi.org/publications/gephi-bastian-feb09.pdf

Jacomy M., Heymann S., Venturini T. , Bastian M. (2012): “ForceAtlas2, A Continuous Graph
Layout Algorithm for Handy Network Visualization.” (draft) http://www.medialab.sciences-
po.fr/publications/Jacomy_Heymann_Venturini-Force_Atlas2.pdf

Noack A. (2009). “Modularity clustering is force-directed layout,” Physical Review E, vol. 79(2).

Thorp J. (2013). Visualization as Process, Not Output. HBR Blog Network, April 3.
http://blogs.hbr.org/cs/2013/04/visualization_as_process.html

Waltman L., van Eck N. J., Noyons E. C.M. (2010). “A unified approach to mapping and clustering
of bibliometric networks”, Arxiv, http://arxiv.org/abs/1006.1032

Clement Levallois