In this paper, we present a platform designed to explore visually massive cooperation between individuals. With the increasing importance of the Internet, new types of cooperation are becoming common, in which hundreds, thousands or millions of individuals act together in interaction, and produces content in a decentralized manner. As these processes are happening in real-time and without organization, individuals involved in them often do not have a clear vision of what is happening, or even which role they play in it. The visualization we propose would allow users to take back the power of understanding the processes to which they participate in. We combine time series visualization, together with custom network visualization, in a way generic enough to adapt to many situations, while o ering numerous possibilities.
Since the advent of the digital era, both the technical possibilities and the introduction of new behaviors have participated in the production of large databases storing tremendous amounts of varied information. Recent hot topics such as Big Data, Complex Systems and Network Analysis have been stimulated by this new access to information. One particular topic of interest is the study of how crowds are involved in massive generation of content, whether it be on Wikipedia, Twitter, Facebook, YouTube, or even through the publication of ever growing number of scienti c publications. If these datasets are a stimulating opportunity, they are also a challenge. While many research has been done on these topics, we feel there is no simple, generic method to explore this decentralized creation of content, and in particular its dynamic. The platform we propose is generic enough to take input from many kinds of sources, such as scienti c publications, online social networks, and many others. The platform is developed with internet based tools only, and could therefore be adapted to provide a user-friendly interface to explore a large dataset of content creation available on the internet.
Several visualizations have been proposed to understand complex systems and large data in general. We introduce the most closely related to our proposition.
ThemeRiver [
History ows [
On a more static perspective, numerous tools, frameworks
and softwares have been proposed to represent networks in
the best possible way. We can cite some of them, among the
most famous ones: Gephi[
The tools we have cited above are either specialized on the visualization of longitudinal aspects, but without information on the internal structure, or, on the contrary, represent this internal structure (network visualization), but only with a static point of view. Our platform is designed to encompass both aspects.
In order to illustrate the possibilities and possible practical applications of the tools presented in this paper, we applied them to three large datasets from di erent elds. In this section, we will present brie y these datasets, and the type of data we extract from them.
For a dataset to be visualized using our platform, it needs to be composed of several productions, that we call Cooperative Productions (CPs). It can be a video, an article, a website, a message, or any other item which can make a reference and be referenced. These CPs are de ned by the following properties:
Category (a chain of character, can be omitted)
Additionally, we need to group these CPs in Cooperation processes. A cooperation process is a set of CPs corresponding to a same topic, a same goal, or any other way of grouping them relevant to the studied dataset. In the following sections, we will detail these properties in 3 example datasets.
NicoNico, or Nico Nico Douga, is a Japanese video-sharing
platform, with functionalities similar to those of YouTube.
With o cially more than 20 Million registered users, and
being ranked among the top 15 most visited websites of Japan,
it is a major Web 2.0 platform. It is especially famous for
the important community of people cooperating in the
creation of complex Music Videos centered on the character of
Hatsune Miku. Starting from an original song, many people
create videos based on it, with innovation such as dancing,
singing, creating new graphics, etc. More information about
this character and phenomenon can be found in [
In NicoNico, tags are associated with videos. We automatically detect tags corresponding to songs with more than 500 related videos. These videos compose the cooperation processes.
Category : extracted from keywords, examples are: Dancing, Singing, 3D, Animation...
References: authors include references to other videos in their comments.
We obtain 165 cooperation processes, composed by 500 to 7654 videos, with an average of 865 videos.
Twitter is one of the most famous and largest Online
Social Networks. In this paper, we consider the di usion of
a particular tweet as our cooperation processes. We used
a dataset covering the period between March 5, 2011 and
March 24, and which covers most tweets published in Japan
during this period. Authors of this dataset claim to have
validated that 80% to 90% of all published tweets appear in
their dataset. For more information, please refer to [
We rst counted for each tweet in our dataset the number of
time they were retweeted, following the method described in
[
Category : Distance in the follower network between original author and retweeter
45 cooperation processes corresponding to retweet chains are
detected, involving between 500 and 2100 tweets, with an
average of 755 tweets.
2.3 DBLP
Massive cooperation predates the apparition of the World
Wide Web. Thousands of researchers around the world
cooperate to improve the global scienti c knowledge. We use
as a dataset the DBLP database [
As we lack topic information, we de ne a cooperation process for each article, with all other papers making a direct reference to it composing the cooperation processes. This de nition is not perfect, but, as we know that seminal papers tend to act as " ags", that must be cited by everyone working on a speci c topic, looking at all papers citing a seminal one is an approximation of a group of works in the same topic. We ltered out all cooperation processes with less than 500 elements.
After ltering, we obtained 41 citation ows, composed of between 500 and 3651 papers, with an average of 664 papers.
The platform we propose is composed of two parts: the time series visualization and the cooperation ow visualization. The time series provides a global understanding of the different cooperation processes studied, together with global indicators on them. In this view, only the global properties are represented, not the individual agents and their interactions. From this global view, it is then possible to select any cooperation process and to visualize its inner details in the cooperation ow view.
In this second tool, in which the details of the cooperation is displayed, several options are possible such as positioning according to time or to step of cooperation, selecting the number of nodes displayed, etc. The navigation between these di erent displays in represented in Fig. 1
When we are interested in a cooperation process, it is
often useful to have rst a global vision of it. We would like
to be able to answer general questions such as: when did
this process started? Is it already nished? Is it becoming
more or less popular? Are there some patterns in its
popularity? These are the global properties of this particular
cooperation.
3.1.1 Macro-level visualization: time series
The visualization we propose excludes the role of each
speci c element, to represents the process as a whole. To do
so, we choose to transform our data in time series, as much
work exists on the topic of time series analysis. For a given
dataset, we de ne a time step, which can be any period of
time (minute, day, year, etc.) and count the number of CPs
published for each category in each time step. For each
Cooperation Flow, we obtain as many time series as there are
categories. We display them as a shape, as shown in Fig.
7. The shape is constructed as a cumulative area chart
augmented with a mirror image of itself, to have a symmetric
shape. The lecture of it is identical to a normal
cumulative area chart. We choose this shape instead of a normal
cumulative area chart because we want to represent
several of these shapes on a same plot with a single time axis.
Therefore, the shape is not framed by the axis, and when
displayed on top of each other, it becomes more natural to
have a horizontally symmetrical shape, as represented on
Fig. 8. A similar observation has been done by the authors
of ThemeRiver [
By displaying several shapes on the same chart, we are able to visually compare them. Examples of interesting observable facts include (but are not limited to): The presence of bursts at a particular location, or following a x period Di erences between cooperation processes starting at di erent times
3.1.2 metrics and graphics
For each cooperative Process, we compute its lifespan, dened as the time between the rst not null value of the time series to the last occurrence of 3 consecutive not null values. This limit is arbitrary, but the objective is to give an end to a time series, potentially in nite, as a new CPs can always occurs in the future. If these 3 non-null values are the last 3 values of the time series, we consider the cooperation process as "still alive". The distribution of the lifespans is displayed as a bar chart.
We compute the normalized centroid of each cooperative ow. The centroid of the time series is the step such as there is as many CPs before and after it. We normalize it by computing:
N ormalizedCentroid = centroidT ime deathT ime
birthT ime birthT ime : A normalized centroid inferior or superior to 0.5 re ect the fact that most of the CPs where produced in the beginning or in the end of the lifespan of the cooperation process. The distribution of the normalized centroid is displayed as a bar chart.
Burst detection is a common problem on time series. A burst
is de ned as a period of time during which the time series
reach temporarily exceptionally high values. In a
cooperation ow, such a burst can typically appears in the beginning
(initial burst), at a given moment, driven by internal events
(new popular CP), or external factors. An interesting case
is when this external factor is not unique but periodic,
typically daily or yearly events. We therefore implemented a
research of such periodic bursts. We implemented the burst
detection with a simple but e ective technique, presented in
[
The relative importance of di erent categories along time We found 5 cooperation processes with periodic bursts in the NicoNico dataset, and we checked that all of them corresponded to yearly events (songs about Christmas, Halloween, etc.).
Whereas the time series visualization allow us to have a quick understanding of global properties, it is often useful to have more insights in the details of what is happening inside each cooperative topic. In this second display, we combine a visualization called cooperation ow together with some alternatives displays and indicators, each of them emphasizing one aspect of the studied cooperative topic.
Original nodes (roots) Level 1
Level 0
Level 2
Time (for a same level)
Levels of cooperation
3.2.1 Cooperation Flow
To represent the details of the process of cooperation, we use
a type of visualization described in [
We observed that one characteristic which can vary greatly between cooperation ows is the importance taken by the most important productions. In some cases, a single production, or a small subset of them, can generate most of the CPs, that is, most of the CPs will directly reference it as a unique source, either during the whole lifetime of the ow, or just during a given period. To study this, we propose a visualization in stacked area of the impact along time of the top 5 nodes, toped by the impact of all remaining nodes (Fig. 6). The lifetime of the ow is split in 10 sections. The impact of a given CP during a given section is computed as the number of CPs published during this period that reference it. We use a black and white scale to avoid confusion with the categories of CPs, already represented by colors. Together with this visualization, we propose a metric to measure this e ect, called CSC, for Cooperation Source Concentration.
CSC =
Pv2T op1 jfu : (u; v) 2 Egj jV j 1 Where T op1 is the set of the 1% nodes with the highest indegree. This metric vary between lim0 and 1, where lim0 is the case where all nodes have the same in-degree, and 1 is reached when all nodes are successors of a single original source. (star-like network) We give the average values of CSC for our 3 datasets in Table 1. We can observe large di erences, with NicoNico having the strongest CSC and Twitter the lowest.
We observed that in some cooperation ow, there is not much cooperation after the rst few levels -the number of CP by level follows a fast shrinking trend- while, in others, it is not the case. This might re ect the ability to renew the
NicoNico 0.92
In this section, we brie y present some examples to show the interest of our visualization. interest in the trend by new CPs. We propose a visualization of this e ect by a stacked bar chart graph (Fig. 6). Each bar represent a level, and we simply count the number of videos of each type published in each level. Together with the general trend, this chart allows to see a change in the categories correlated with the level. The color used for the categories are coherent with the ones used in the cooperation ow chart.
The indicator we propose to summarize this chart is SC, Sustainability of Cooperation. It is de ned as the average of the variations of the number of CPs between successive levels, pondered by the number of CPs in the rst of the two:
SC =
Pnl 1 nbCP (i+1) i=1 nbCP (i)
(nbCP (i + 1) + nbCP (i)) nbCP (1) + 2 Pin=l 2 2 nbCP (i) + nbCP (nl 1) with nbCP (i) the number of CPs at level i, and nl the number of levels. SC=0 if there is no production after the rst level. SC >1 if the number of CPs tends to grow with each level. The lower the SC value, the less CPs tend to generate new cooperation. In Table 2, we represent the average value of SC for our datasets.
NicoNico 0.23 Twitter 0.39 DBLP 0.44
4.2.1 Deep study of one dataset: NicoNico NicoNico is the richest and the most complex of our datasets. In g. 9, we show 2 typical ow from this network. We can make the following observations, also valid on most other ows: 1. There is only one original source, and most of the cooperation is made directly from this source, as we can judge by the large RI-Torus 2. Most important nodes for the collaboration are on the rst level, they directly reference the original node only 3. The cooperation is more wide than deep, there is not much cooperation at a level greater than 3. 4. Although many categories (colors) are present, each node seems to generate a specialized cooperation: RITorus are mostly of a single color, not always the same. 5. There is no strong correlation between the number of view of a video (area of inner circle) and its capacity to generate cooperative behavior (torus area) 4.2.2 Comparison of datasets In g. 10, we present two visualizations typical of the other datasets. We can immediately spot some di erences. In the tweet dataset, cooperation is deeper, and we tend to see the formation of chains, long but without many bifurcations. More important nodes are not necessarily situated at the rst step, but can occur deeper. There seems to be a stronger relation between the popularity of the node and its role in the cooperation. There is not a single source. In the citation dataset, we immediately spot a large number of nodes making references to several others. These nodes with many references are important in the cooperation. Nodes at a deep level seem to generate as much cooperation as those in the rst levels. There also seems to be a lesser concentration in the cooperation generation: a larger fraction of nodes are referenced by other important nodes, and the gap is less important between the top in uential nodes and the ordinary ones. Exploring in further details the properties of the di erent datasets is beyond the scope of this paper.
In this paper, we have presented a platform to explore mass cooperation, and a set of tools to explore di erent aspects of this type of cooperation. Our conception of such visualization was driven by our previous experiences in the exploration of large datasets formed by cooperation, and the di culties encountered to understand the underlying mechanisms.
We also presented some complementary visualizations and metrics that focus on several aspects of the data, with different granularities, and can also help to apprehend it. In the future, we hope that other researchers will use this platform and help to improve it, either by their remarks or extending the possibilities. In this prospect, we release its source code, altogether with an interactive online version, so as interested researchers could work with it as easily as possible. In particular, it could be interesting to add metrics and statistics, such as a one could choose the more interesting indicators in his case. The source code and browsable example is available on the website of the rst author. Another future possibility is to propose Internet applications based on this visualization to the destination of nal end users. For example, one can think of a plug-in for Google Scholar allowing to browse research topics.
We thank Fujio Toriumi for collecting the Twitter dataset, and allowing us to make use of it in this work.