Content uploaded by Weimao Ke
Author content
All content in this area was uploaded by Weimao Ke
Content may be subject to copyright.
Movies and Actors: Mapping the Internet Movie Database
Bruce W. Herr, Weimao Ke, Elisha Hardy & Katy Börner
School of Library and Information Science, Indiana University, Bloomington, IN 47405
{bherr@indiana.edu, wke@indiana.edu, efhardy@indiana.edu, katy@indiana.edu}
Abstract
This paper presents the results of an analysis and
visualization of 428,440 movies from the Internet Movie
Database (IMDb) provided for the Graph Drawing 2005
contest. Simple statistics are presented as well as a
tapestry of all movies with an overlay of the giant
component of the co-actor network. Academy award
winners are highlighted. Major insights are discussed.
Keywords---network analysis, domain visualization,
movies
1. Introduction
Since 2002, the International Sunbelt Social
Network Conference has hosted a so called Viszards
session [6] that aims to show the power of network
analysis and visualization. The work discussed in this
paper was done for Viszards 2006 at Sunbelt XXVI
which took place in Vancouver, BC, Canada on April
28
th
, 2006. Viszards 2006 asked network science
researchers to analyze data retrieved from the Internet
Movie Database (IMDb). IMDb (http://www.imdb.com)
is a popular site cataloging almost every movie ever
made.
The study of IMDb data is interesting for several
reasons. For one, most people know about and can relate
to movies and actors. Thus, when presented with a
visualization of movie data, they will try to find their
favorite movies and actors, identify movies of potential
interest or explore the complex co-actor relationships
among actors. Second, the dataset has rich information
on each movie and actor allowing for a wide variety of
data analyses. Third, the dataset is sufficiently clean and
structured so that analysis can be done without using
semantic matching techniques.
From the beginning, our goal was to show all
movies as well as major co-actor relationships. We
wanted to give the world an overview of the movie and
actor space that almost everyone is familiar with. Doing
this on a large canvas (the final visualization has a size of
36” high and 73” wide) and in a way that people can
reason about and understand the visualization was a
major challenge. The required data density due to data
volume per square inch posed additional difficulties.
With this paper and the IMDb visualization we hope
to communicate the power of visually pleasing yet
informative visualizations to a general audience.
Visualizations can be more than eye candy. Paper
printouts are discussed as a viable alternative for the
presentation of high density visualizations.
The remainder of the paper is organized as follows:
Section 2 introduces the dataset used. Section 3 explains
the data analyses and results. Section 4 discusses the
iterative design of the visualization and insights gained.
The paper concludes with a discussion and outlook.
2. Data preparation
The data for the IMDb visualization originates from
the Graph Drawing 2005 web site [3] at
http://www.ul.ie/gd2005/dataset.html. The dataset is a
bipartite graph in which each node either corresponds to
an actor or to a movie. Edges go from a movie to each
actor in the movie. It also provides metadata for the
nodes like movie/actor name, year of the movie, and
genre of the movie. This data was then parsed and stored
in a relational database to ease data manipulation.
As with all large datasets, there were diverse
anomalies. Out of the 428,440 movies in the set, 2,091
movies had no year data, six movies were produced in 1
CE, two were produced in 2 CE, 24 more were produced
between the years 3 and 1888 CE, and the ‘Adult’ movie
entitled ‘Westside Boys’ is to be produced in 9006 CE.
The biggest anomaly in the derived data is the fact that of
the 428,440 movies provided, 123,617 movies have no
actor data at all. This is particularly problematic for us
since we are showing the interplay between actors and
movies. We believe that this is most likely a problem
inherited from the derived data, since the official IMDb
statistics say that as of March 2007 (the derived data was
from early 2005) there are 365,328 movies in the
database. In the end, we excluded those movies that did
not have actor information.
Herr II, Bruce W., Ke, Weimao, Hardy, Elisha, and Börner, Katy. (2007) Movies and Actors: Mapping the Internet Movie Database.
In Conference Proceedings of 11th Annual Information Visualization International Conference (IV 2007),
Zurich, Switzerland, July 4-6, pp. 465-469, IEEE Computer Society Conference Publishing Services.
3. Data analysis
After getting the data into a relational database,
several statistics were run to get a feel for the data. We
excluded the anomalous data discussed in the last section
resulting in 302,691 movies produced between 1890 and
2007. It should be noted that this data was from early
2005, so all movies beyond that were in differing stages
of production and not yet released. Figure 1 shows the
growth of movies over time. The red lines show the
boundaries of the movies that we considered.
Figure 1. Growth of Movies Over Time
Figure 2. Movie Out-Degree Distribution
There are 3,792,390 links connecting 302,691
movies and their 896,308 actors, see Figure 2. 38,027
movies have exactly one actor. Movies with more than
1,000 actors are ‘The Eurovision Song Contest’ (1,338
actors), ‘Around the World in 80 days’ (1,287 actors),
and ‘General Hospital’ (1,123 actors).
In order to get a feel for the actor space, we created
a co-actor network where actors are connected based on
the movies they acted together in. Actors that appear in a
movie together are said to co-act. The network of co-
acting actors contains 896,308 actor nodes and
114,128,535 co-actor links, see Figure 3. Each link is
weighted by the number of movies the two actors were in
together.
Figure 3. Co-Actor Out-Degree Distribution
4. Data visualization
Our main goal with this visualization was to give a
global overview of the entire movie and actor space.
During the initial design phase, we wanted to draw a co-
actor network and have it be surrounded by a list of all
movies. However, fitting this into a reasonably sized
canvas proved difficult. The sheer number of movies to
plot was so large that we eventually decided to render
them in columns and overlay the co-actor network
directly on top of the movies. We also wanted to
constrain ourselves to 36” high and somewhere around
40” wide, but eventually went with 73” wide to
accommodate all of the movies. A description of the
layers of the final visualization follows.
At the bottom of the visualization is the movies
layer. The movies were grouped by year and plotted in
97 columns. Within each year, the movies were sorted
and their titles size encoded by the number of starring
actors. Furthermore, movie titles are sorted and color
coded by genre. Each of the seven top genres (Short,
Drama, Comedy, Documentary, Adult, Romance, and
Thriller) was given a distinct color while the rest were
given a light grey color. Over plotting was utilized to fit
the movies into the area provided. A white outline is
drawn around each character to improve text legibility. A
close-up of the movies layer is given in Figure 4.
Figure 4. Zoomed View of the Movie Layer
The next layer up is the actor layer. We felt that the
best way of showing the actors was by laying out their
co-actor network using a force-directed layout algorithm.
Each edge between the actors was weighted by the
number of times the two actors had been in the same
movie together. Interested to see the strongest co-
Herr II, Bruce W., Ke, Weimao, Hardy, Elisha, and Börner, Katy. (2007) Movies and Actors: Mapping the Internet Movie Database.
In Conference Proceedings of 11th Annual Information Visualization International Conference (IV 2007),
Zurich, Switzerland, July 4-6, pp. 465-469, IEEE Computer Society Conference Publishing Services.
actorship linkages we excluded all those links that had a
weight of less than three. Resulting unconnected nodes
were excluded. The remaining core network with
105,758 nodes and 1,292,816 edges was fed into VxOrd
[2] to lay out actor nodes with a modified spring force
layout algorithm. This algorithm ensures that highly
interlinked nodes are close to each other and unlinked or
weakly linked nodes are further apart. The resulting list
of coordinates for each of the 105,758 actor nodes was
rendered using Pajek [4]. The color of each actor node
corresponds to the movie genre s/he most contributed to.
The results are shown below in Figure 5. A zoomed in
portion of the co-actor network can be seen in Figure 6.
Figure 5. Co-Actor Network
Figure 6. Zoomed View of the Co-Actor Network
Another layer was added to provide landmarks in
this complex co-actor network. The network was
spatially cut into a 10x10 grid and the actor node that had
been in the most movies in each of the cells was labeled
with their actor name using a light colored, 15-point font.
These labels are useful in identifying clusters of actors.
The discussed layers form a reference system that
can be used to overlay additional data. In the
visualization described in this paper, we added two more
data layers. The first shows Academy Award’s best actor
and actress winners and nominees from 2000-2004 [1].
They are represented as 41 darker and larger actor labels
on top of the co-actor network. The most interesting part
about this layer is that most of the actors are tightly
packed into one cluster. Though not fully explored, this
may mean that in order to increase one’s chances of an
academy award for best actor/actress, one should work
closely with actors in this cluster.
The second additional layer dealt with the winners
and nominees for the Academy Award’s best picture
award. The 25 movies nominated (including the winners)
have exactly 433 actors in the co-actor network. This
layer draws lines from the 25 nominated movies in the
underlying movie layer to the associated actors in the co-
actor network layer. The color of the lines corresponds to
the genre of the movie. The curves of the lines were
chosen so as to not cover up too much of the co-actor
network. This layer helps to highlight what areas of the
actor space is being used by top movies in the field.
All of the layers except for the co-actor network
were created with custom code that reads in the provided
data and produces PostScript® files. The co-actor
network’s layer was outputted to PostScript through the
Pajek program. To produce the final image, the assorted
layers were combined and rasterized at 400 dots per inch
(DPI) in Adobe Photoshop©. An additional layer was
created in Photoshop that added the informational
column on the right side of the visualization.
The movies layer proved to be very difficult to
rasterize in Photoshop due to its size and complexity. For
the version presented at Sunbelt, we had to reduce its
complexity by removing the textual outline drawing.
This worked, but we were never quite satisfied with the
loss in quality that resulted. After nine months of trying
larger machines and distributed rendering, a solution was
found. By using the GNU Image Manipulation Program
(GIMP) and utilizing a Sun server with 32 GB of RAM
and 4 processing cores, we finally got the layer to
satisfactorily render at 400 DPI. This new image has
replaced the older movies layer.
Figure 7 at the end of this paper shows the final
visualization. Unfortunately, it is more than eight times
smaller than the original visualization and many details
are lost at this size. To really appreciate the visualization,
one must either have a full resolution printed version or
go to http://scimaps.org/maps/movieactors to see a
zoomable Google Maps interface to the visualization.
The map is also available for sale from
http://scimaps.org/ordermaps in support of the Places &
Spaces: Mapping Science exhibit.
Discussion
The presented work demonstrates the utility of paper
printouts for serving high data density visualizations.
Paper as a medium is easy to access and transport, offers
high data density, and is comparatively cheap. Humans
have used paper and interacted with it for well over
2,000 years and have highly optimized it as a medium to
store, transmit, and preserve information. Paper naturally
supports exploration. Interactivity like zooming and
panning can be accomplished by physically moving
closer to and further away from the print. While there are
problems with zooming and panning in computational
environments, this sort of interaction with paper is
immediately obvious to viewers. Arbitrary annotations
are possible. Last but not least, there is something to be
Herr II, Bruce W., Ke, Weimao, Hardy, Elisha, and Börner, Katy. (2007) Movies and Actors: Mapping the Internet Movie Database.
In Conference Proceedings of 11th Annual Information Visualization International Conference (IV 2007),
Zurich, Switzerland, July 4-6, pp. 465-469, IEEE Computer Society Conference Publishing Services.
said about a visualization that can be physically touched
and has a real texture to it.
The higher density of paper has allowed us to give
an overview of the entire movie and actor space in a
reasonable physical space. Our current visualization
renders at 400 DPI, but there are techniques to utilize up
to 4,000 DPI. The result is extremely crisp graphics that
allow for further zooming with a physical magnifying
glass.
In addition to bringing out paper’s natural strength,
this visualization work also made obvious the current
limitations of rendering on large display walls. Display
walls are limited by the rather low resolution of modern
monitors and projectors. To render the full resolution
visualization, a display wall would have to be around 4
times as large as the equivalent print. A 12’x24’ display
wall would be prohibitively expensive. Compare this to
an equivalent 3’x 6’ print which is much cheaper, denser,
and could be mass produced.
Future work aims to update the data behind the
visualization and add a layer of interactivity. We will do
this by utilizing an invention by W. Bradford Paley
called an illuminated diagram [5]. This technique uses a
projector to interactively highlight interesting parts on
statically printed diagrams. We can then take advantage
of the interactivity of computers, yet still retain the
qualities of printed media.
Acknowledgements
We would like to thank all those involved in the
Internet Movie Database for creating an excellent
dataset, Vladimir Batagelj for organizing the Viszards
session, Bryan J. Hook for editing, and Sumeet Ambre
for creating the Google Maps interface for our
visualization.
This research is supported by the National Science
Foundation under IIS-0513650 and a CAREER grant
under IIS-0238261. Any opinions, findings, and
conclusions or recommendations expressed in this
material are those of the author(s) and do not necessarily
reflect the views of the NSF.
References
[1] Academy Award’s best actor and actress winners and
nominees from 2000-2004 downloaded from
http://www.imdb.com/Sections/Awards. Accessed on
April 2006.
[2] Davidson, G.S., Wylie, B.N. and Boyack, K.W. Cluster
stability and the use of noise in interpretation of
clustering. Proc. IEEE Information Visualization 2001.
23-30.
[3] Internet Movie Database (IMDb) network provided for
GD’05 at http://www.ul.ie/gd2005.
[4] Nooy, W.d., Mrvar, A. and Batagelj, V. Exploratory
Social Network Analysis with Pajek. Cambridge
University Press, 2005.
[5] Paley, W. Bradford. Illuminated Diagrams: Using Light
and Print to Comparative Advantage. InfoVis Conference
2002.
[6] Viszards: Analysis and Visualization of IMDB Networks
session description from
http://www.insna.org/2006/special.sessions.html.
Accessed on April 2006.
Herr II, Bruce W., Ke, Weimao, Hardy, Elisha, and Börner, Katy. (2007) Movies and Actors: Mapping the Internet Movie Database.
In Conference Proceedings of 11th Annual Information Visualization International Conference (IV 2007),
Zurich, Switzerland, July 4-6, pp. 465-469, IEEE Computer Society Conference Publishing Services.
Figure 7. Complete Map
Herr II, Bruce W., Ke, Weimao, Hardy, Elisha, and Börner, Katy. (2007) Movies and Actors: Mapping the Internet Movie Database.
In Conference Proceedings of 11th Annual Information Visualization International Conference (IV 2007),
Zurich, Switzerland, July 4-6, pp. 465-469, IEEE Computer Society Conference Publishing Services.
