Thinking in networks:
what it means for policy makers
Alberto Cottica, University of Alicante – RENA Summer School 2013
Hello, thank you all for showing up. I do research for a newly minted company called Edgeryders. The reason why the
company even exists and why its business model looks the way it does itself has a lot to do with what we are going to be
talking about today. But more on that later.
•Thinking in networks
•Making policy in a
world of networks
Act 1: Thinking in networks
Research Institute of Molecular Pathology
A network represents relationships across entities
Networks are mathematical objects. They provide us with a formalized way to think about relationships
(represented by edges) across entities (represented by nodes). They are very general: you can use them to
describe relationships of any kind across entities of any kind. And people do: they study networks of genes, where
nodes represent genes and edges represent the same disease being encoded by two genes. Food chain networks,
where two species (nodes) are connected by an edge when one feeds on the other. Ingredients in cooking, where
two ingredients (nodes) are connected by an edge if they appear in one dish together. Blogging networks, where a
blog (node) is connected by an edge to another if it links to it. And on it goes. There are airline networks to study
logistics; power grid networks to understand baseload and cascading failures; both the Internet and the World
Wide Web have well-studied network representations; ﬁnancial networks to understand contagion from a few
struggling banks to the rest of the ﬁnancial system; there’s even studies on networks of dolphins swimming
together. A particularly important type of network for policy makers are social networks: these are networks in
which nodes are people.
Behavioral change propagates by “contagion”
Social networks are a useful way to think about societies, the economies they support and policies enacted on
them. Why? Well, because most policies are about affecting the behavior of agents one way or another. And it
turns out behavioral patterns travel across social links. Among the early adopters of network models are
epidemiologists. Think, say, of AIDS. You model your population as a social network: two people are connected if
they are sexual partners. If you know who is a partner of whom, and the probability of being infected by sexual
contacts, you can predict the pattern of the epidemics.
So far so good. But here’s the curveball: someone used the same model for obesity, and got a really good ﬁt. If
you have obese friends, your probability of being obese is signiﬁcantly higher. Why is that? There is no obesity
virus or bacterium that you can transmit through social contact. So they tried contagion models on other
theoretically non-contagious states. Smoking: good ﬁt. Giving up smoking: good ﬁt. Income: good ﬁt. Getting
divorced: good ﬁt. Unemployment: good ﬁt. You get the idea. It looks like we are wired for imitation and
sensitivity to social pressure, for good or bad. Nicholas Christakis’s TED talk is an entertaining introduction to this
See and reuse the social infrastructure
So, thinking in networks means you train yourself to see social networks – patterns of social interaction across
individuals – as a fundamental communication and coordination infrastructure underpinning society. Once you
know it’s there, you start thinking about policies in fundamentally different way: policy becomes a signal that
travels through that network, whereas now we think of policies as a broadcast signal. Suppose you want to get a
message from A to B. In a broadcast world, you just broadcast. It’s a brute force approach: you just push out the
signal in all directions, hoping there is no wall to block your signal.
But if you are aware of the network that points A and B are a part of, you can get your message delivered in a
much more efficient way. Suppose that point A is Dundalk and point B is Athlone: if you know the road network
you’ll go through Dublin rather than walking across the countryside in a straight line.
With networks, it is surprisingly easy to get interesting behavior out of very simple models. You are looking at a
model of preferential attachment. Just assume a networks grows. Nodes come online and they have one edge to
invest. They will connect to another node at random, but the probability of connecting to any node will depend on
the number of links that the node already has. Which makes sense: suppose this is a social network, you might
care about making sure you are friends with the person that know everybody, and might give you access to more
people. Granovetter has shown decades ago that people ﬁnd jobs through weak social ties, acquaintances rather
than family or friends. This simple model generates a network structure that mimics very well that of real-life
social networks. Thinking in networks trains you to look at highly organized social system without necessarily
postulating a social planner, or a leader or something like that. By implication, it teaches you humility: you learn
that peer-to-peer social interaction, left to its own devices, can generate sophisticated structure. Sometimes you
might not need policy at all: just leave things alone, and let the network dynamics work their magic.
This process also sheds light on the much-debated inequality. Preferential attachment leads to a distribution of
links very, very unequal, described mathematically as power laws. It means the number of nodes having n links is
a negative function of e to the n. The famous “1% of the world’s population own X% of global wealth” is a way to
express a distribution of wealth governed by power laws. There are two consequences of this: (1) representative
agents lose signiﬁcance: almost no one is “average”, everybody is poor except for a few very rich people, so
designing a policy for the average income people is probably going not to work for anyone. (2) Power laws are
found everywhere in natural sciences, and they typically signal complex systems at work. Inequalities might be a
structural feature of complex systems – including human societies. Yesterday Stefano and others described how
participatory budgeting in Porto Alegre bounced back from an equal access situation at start towards a more or
less ﬁrst-world looking stakeholder system.
Impact: the right tool for the job
Impact is an obvious one. You get more bang for your buck: you are trying to ﬁght aids by focusing on the few
network hubs, people with very many sexual partners, and going graphic on them rather than putting up vague
posters in schools and community centers.
Iatrogenics: harm done by the healer
As our societies get ever more complex, they get ever more difficult to second guess. There is a real risk of what
Nassim Taleb calls iatrogenics, harm done by the healer.
2007-2013 – Billion €
World Bank lending commitments
Italy, strategic national framework pipeline
One of my favorite examples of that is with public spending. In my country, Italy, we have a situation. The north of
the country is well-developed, with quite a strong manufacturing economy, whereas its south is lagging behind.
This is a high political priority, and for at least ﬁfty years we have thrown money and brains at it.
2007-2013 – per capita €
World Bank lending commitments
Italy, strategic national framework pipeline
That means people in that part of Italy, per capita, see a substantial pot of money – 200 times their counterparts
in the rest of the world.
“Everyone was talking about
public sector tenders.”
– Tiago Dias Miranda in southern Italy, 2013
The result of this situation: smart, entrepreneurial young people in Italy’s Mezzogiorno are talking about public
sector tenders. They know all the acronyms of European programs. And why not? Though most of the money ends
up with networks of incumbents, even the crumbs can be quite a big payoff. But of course, in development terms,
this is just a distraction: as they write funding applications, they are not starting companies, or leaving the
country, or squatting buildings; they are not engaging in collective, trial-and-error discovery of the paths that
lead to the healing of the economy. And sure enough, the economy does not heal. The government means mostly
well, but the amount of damage inﬂicted is terrifying. Thinking in networks helps in two ways: ﬁrst, it teaches you
a healthy respect for self-organizing social phenomena; second, it encourages deploying narrowcast, minimal
intervention rather than broadcast heavy ones.
“Too big to know”
In the age of big data, it’s paradoxically getting increasingly difficult to take responsibility for decisions made on
the basis of evidence. Why? Because evidence is difficult to interpret. Take machine learning: we get our result by
evolving algorithms to make decisions, then feed them unfathomable quantities of data that you can’t possibly
inspect visually. Even the people who trained the algorithms have trouble interpreting what they do: for most
senior decision makers it is unrealistic to take courses in linear algebra and data science just to hack apart a
result. In early 2013 it turned out that an inﬂuential paper by Carmen Reinhart and Kenneth Rogoff, observing a
tendency to sovereign default in states with a debt/GDP ratio over 90%, had an Excel error. Network modeling is
relatively intuitive, in the sense that you can get quite far on simple, intuitive models.
Compassion: it’s not you, it’s system dynamics
Remember the preferential attachment model? We simulated the existence of superstars starting from identical
nodes. Superstars are desirable in many network, because they result in a topology called scale-free. One of its
properties is that propagating information across the network can be done very efficiently thanks to the “hubs”
with many links, connecting everyone to everyone else. But this efficiency property at the system level comes at a
price: high inequality at the individual level. And this inequality seems unfair: most superstars acquire their status
by being born early, or getting a lucky break early on. The system dynamics does the rest.
Most network models assume identical nodes: network math makes you very aware that your special position in
society can be explained as a function of variables you have no control on, and, in many cases, of plain
randomness. A node gets a “lucky break” early on, in the sense of getting two or three connections, and system
dynamics propels it into superstardom. Because of well documented psychological biases (humans like coherent
stories and did not evolve the ability to process well randomness), superstars will rationalize ex post their
privileged position as the result of talent, hard work, or the favor of God. Thinking in networks challenges the
“underserving poor” rhetoric and leads to empathy for the people who get pushed to the left of the degree
distribution, that might well be as smart as some of the superstars or better.
Measure: quantify online social interaction
Social media are a game changer in this space. Because of the technology we use to support it, online social
interaction leaves traces encoded in databases. You can then mine those databases to rebuild the graph of social
interaction. This is what Google and Facebook are doing. They care about the contagion dimension of
consumption, and microtargeting, say, luxury watches ads to the people with the highest probability of buying
luxury watches based on what their friends do. The logic is “hey, your friend has bought XY! What are you going to
do?” (by the way, compare the cost-effectiveness of this with that of a TV campaign on AIDS prevention). But we
can use them to map the transmission of other behavioral change signals – and possibly to inﬂuence it.
This opens up interesting scenarios. In my own work, I study online conversation in participatory processes, and
try to ﬁgure out how to diagnose its health by looking at the shape of the interaction networks. Our ultimate goal
is to drive participation processes – where, by deﬁnition, we have zero command and control over individual
participants – by taking advantage of the inﬂuence on individual participants of the global networks characteristics.
On these, whoever is running the participatory exercise typically does have some control: for example, you can
make the network more dense by exposing participants to a feed of content generated by people they are not
Act 2: Exploring networks of conversation
This brings me to the second part of this lecture, in which I zoom into what it means to explore an actual network
of people participating in a collaborative policy exercise.
The early days
In the early days of network science (1960s: then called “sociometrics”), researchers would specify social networks manually.
They would go around and ask people about their behavior or their feelings. Which families does your families visit in the
village? Which of your schoolmates would you rather sit at the table which at lunchtime? Who is married to whom? You’d get very
simple graphs like this one. Clever researchers like Padgett and Ansell would then try to get mileage out of these small sets of
data. Their famous study on the network of marriages across the wealthy families in 15th century Florence attempted to explain
the later rise of the Medicis as the dominant family in terms of their high centrality in this network. The implication is that
alliances among Florentine families were greatly facilitated (or signaled: causality is not implied here) by family members being
married with each others. Central families were better positioned to build alliances to isolate their enemies. The scientiﬁc
programme behind these studies was inﬂuenced by the work of Georg Simmel, and carries the promise of explaining society by
something called social structure. Simplifying quite a bit, interaction between social actors cristallyses into metastable
structures. By mapping out today’s interaction patterns, you could in principle predict tomorrow’s societal structure. This was
exciting stuff for a small number of mathematically inclined sociologists in those days, but it was empirically almost irrelevant
because live in very large social networks, not in small ones. And large social networks are impossibily expensive to map by
hand, even notwithstanding the small detail that people lie when answering to surveys, and so you can’t really take survey data
Network data for free
Online interaction environments changed this game. For technical reasons, everything that happens online has to
be encoded in a piece of software called a database. When you write a comment to your friend’s Facebook status
update, Facebook appends a line to a very large table of comments with billions of lines. That line is divided into
ﬁelds, and keeps track of who wrote that comment, when, to which status update and what its content is. The
consequence is this: if you can build a functioning platform for online interaction, you get its monitoring for free,
or for very cheap. The NSA and other more or less secret police bodies know this, and are milking the Internet for
intelligence data. I will not speak about this, since Smari here is arguably one of the world’s leading experts on the
matter. Talk to him if you are curious.
Edgeryders: a case study
In 2011 I was with the Council of Europe and we were tasked with making proposal for the reform of European
youth policies to member states and the European Commission. We decided to roll out a radically open, many-
many participation exercise. People would interact on an online platform and generate the recommendations on
which that policy document could be built.
in December 2012
Average distance 2.3
Along the course of our work, we generated a lot of content by about 250 active participants). We decided to take
advantage of the information in the Edgeryders database to assess the validity of that content. We started by
specifying a social network. Nodes were of course registered users on the platform. Edges were comment: Alice
was connected to Bob if she had commented at least one post or comment by Bob. The weight of the edge is
equal to the number of comments she has written to Bob’s content. So, this gives rise to a network that is directed
(Alice can be connected to Bob without Bob being connected to Alice) and weighted (being connected is not a yes/
no variable, but an integer variable). Here color represents degree: nodes with only few edges are colored green,
nodes with lots of edges are shaded blue, intermediate ones are shaded red.
Once we had that network we discovered interesting things. For example, we were working with a small team that
would help animate the network (about 6 people at any given time, with some turnover) Look what happens when
I remove the nodes representing the team:
Average distance 3.3
What do you conclude?
Also, subcommunities had formed. Once you remove the moderators, this network is very
nonrandom, in the sense that it partitions very well in relatively separated subcomponents,
with subcomponent “leaders”, or inﬂuential people, very visibile. We ended up using this
ﬁgure as a recruitment tool!
Blue edge = more
specialization on one or
few topics, red edge =
no specialization, all
topics equally discussed
Modern-day network analysis can get quite sophisticated. We attempted to look for
specialization across subcommunities (this was possible because our conversation run in
“campaigns”: work and employment, education and learning, politics and participation etc.).
We use vector cosine distance to represent how much the distribution of comments in each
subcommunity deviates from the uniform distribution: but that did not go well, our data were
How “broad” and
“deep” are discussions?
We also could look at discussion topics. Research questions are blue, proposed contributions
green, comments red. This visualization tells you which questions were widely addressed and
which solutions were most commented. If a question receives many candidate solutions and
each candidate solution is widely debated, you can probably trust the outcome of the process
more. Similar techniques are used in the literature to assess the credibility of Wikipedia
Topic tree (detail)
A nice fractal pattern
subtopics, posts and
Why am I telling you all this? I wish to encourage you to try to be rigorous, to distrust your
intuition and make a beeline for the data whenever data are available – which is surprisingly
often. A lot of you care about participatory processes: I do too, and I think they deserve to be
treated with the most rigorous tools available.
Act 3: Making policy in swarms
In recent years, lowering costs of coordination have allowed loose, unorthodox constellations of people in action to achieve
spectacular things. StackExchange and Wikipedia come to mind; more recently, and closer to the politics/public policy world
that we care about, Internet enabled swarms have started attack the dominance of public institutions and traditional
stakeholders on public decision making. You may remember the anti-ACTA movement that spreading from Poland, managed
to overturn a EU-USA agreement very strongly sponsored by just about any important stakeholder. Or the Pirate Party, that
started in Sweden and collected 9% of votes (30% in the under-30 demographics) in a few months; Smari here can tell you
about how he and his co-founders took the Icelandic Pirate Party into Parliament (5% +) in nine months.
At the heart of the “swarm” concept there is a fundamental paradox. Swarms derive their uncanny efﬁciency from radical
decentralization of decision making and action; yet, decentralization might and does cause such action to develop in directions
so different from what it had been intended to be as to be unrecognizable. I guess most of us will be turning around this
paradox in their head. The main tool I am using to debunk this paradox is network theory: I conceptualize swarms as people in
networks. In networks, nodes might be equal in the amount of top-down power over others, but they will typically be very
unequal in terms of connectivity, hence the ability to spread information (including narratives and calls to action) across the
network. Uneven connectivity adds some directionality to the swarm, in the sense that the most connected people get it to go
their way most of the times.
I am a policy guy – public policy design (and some deployment) is what paid my bills for the last ten years. Public policy is
generally understood as a top-down process: some leader somewhere makes a decision and that decision is enacted. Since
the accepted modi operandi of public policy are encoded into law, such top-down thinking is hardwired into organizational
charts, remits and procedures. A decision maker wanting to do things differently will not in generally be enough for things to
happen differently. Think of this as an especially hard area to do swarms in. That’s not a bad thing for today’s purposes,
because it provides us with a clean benchmark. If you can do it in the government, you can probably do it in most places.
All this is very tentative. I can’t claim I know how to do this stuff. I mean, I do it, and it kind of works, but I am not sure exactly
why, so I would be the ﬁrst to not want to turn the revenue agency into a swarm just yet. In fact, the reason why I am here is
that I hope you guys can help me make some progress. I am also going to assume you guys have been thinking into it as hard
as I have, so I am giving you the full complexity of the argument. Stop me if I touch on something that does not make sense to
you, or that you don’t know about.
Let me take you back to iatrogenics. Remember? The government means mostly well, but the amount of damage
inﬂicted is terrifying. This is why I and others are exploring other ways. Iatrogenics in run-of-the-mill public
policy is a powerful argument for exploring the way of smart crowds, or swarms. If you are a citizen seeking to
drive change, it is not a bad thing to explore: you are low on the public policy food chain, swarms give you an
alternative power base, which explains the success of outsiders like the anti-ACTA movement and the Swedish
Public policies as a buyer’s market
But doing policy in swarms has an immediate consequence: you need to recruit people, and those people do not
work for you, do not take (much) money from you and need to be convinced.
... and that’s a big reality check right there. I believe this has given some competitive edge to my own projects. I
just had to work harder to get ANYTHING off the ground.
Timing: get friends to start the bandwagon
Scholars of swarms, social networks etc. focus typically on the behavior of the formed swarm. But if you’ve ever
tried it, you know that the hardest part is to kickstart one. We need a much better developed embryology of
swarms. Me, the better method I know is still to leverage trust network of friends. This is how Vinay jumpstarted
Big Picture Days: he wrote an email to twenty people trying to get the ﬁrst, say, six to commit. Then, he could tell
everyone “You don’t want to miss this cool event. Why, Alberto Cottica is coming!”. Even if you don’t know who
the hell Alberto Cottica is, such a call works with the deep wiring of human psychology. We have plenty of
experimental psychology results around that by now.
Randomness: shake things up (hence parties)
Photo: Medhin Paolos
You are making policy because someone perceives a situation that is not ﬁxing itself. Rather than going in with a
heavy intervention (traditional economists will “maximize the welfare function” and push the economy towards the
maximum), which is iatrogenetic, you can simply shake things up a little bit to see if the system gets unstuck from
its present undesirable attractor and starts moving towards a better one. Complexity thinking has given us,
among other things, an attractive theory of innovation based on generative relationships: innovation stems from
people being similar enough that they can communicate well, but different enough to give each other mild
cognitive shocks, inducing new ways to look upon things. It is not hard to assess the generative potential of a
relationship, but it impossible to predict in advance which potentially generative relationships will actually lead to
So, I just like to throw parties and introduce to each other people from diverse walks of life. Curated parties
increase the number of new connections in your network and therefore, in probability, the number of new things
being tried. This, in turn, increases the probability of your situation unmooring from where it had been stuck. And
no iatrogenics. Win!
Transparency: requests for comments
Photo: Elena Trombetta
I ﬁnd a radically transparent behavior to be advantageous when running a swarm: in a buyers market you need to
win trust. Transparency also doubles up as a management tool: most people will just appreciate that you are
being honest about, for example, how much money you spend and on what, but occasionally somebody pays
close attention and ends up making useful suggestions. If you have to ﬁght a narrative of public policy as corrupt
and self-referential (I do) transparency is an amazingly effective tool in reducing conﬂict and suspicion.
Time bombs: zero entrenchment
Many swarms tend to lose their magic after a while – the mavericks of the early days get suitiﬁed, their project
becomes a job or what have you. I like to build time bombs in my projects: if a swarm is active enough, it will ﬁnd
a way to survive it. In fact my company, Edgeryders, formed with the intention of providing a new core to a
community that assembled around a public sector project I used to direct. The project ended, but some of us felt
the community was too good to pass on, so we decided to build a small organization to provide it with the
scaffolding initially provided by the public sector project.
Efﬁciency: don’t touch the wicked problem
Photo: Alberto Cottica
When you are doing stuff with a swarm and it appears to be working, outside people will try to get it onto
problems they care about. I try to resist this. It implies a revision of the social contract, which tends to be
conﬂictual: also, it might destroy that feeling of effortless impact that core community members ﬁnd intoxicating.
Generally, bad idea. So, if you are running an effective open data community and people try to get you to point
that swarm in the direction of, say, salvaging a badly designed competition for developers to use open data in a
second-tier town, my advice is: say no. Let them fail.
Trust: no strings attached (even give people cash)
Control is costly and boring. Relinquishing it, and rather focusing on enabling people to take initiative makes you
save a lot of time and money, and is a huge motivator, as people feel empowered and trusted. If you can, you
should give people a little cash with no strings attached. There is a recent Ugandan study that provides evidence
that, even without swarms, even giving money to young rural poor results in increased hours worked and
increased income for the people in question.
Measure: do you have enough complexity?
Photo: Alastari Montgomery
To do this stuff, you need a minimum of complexity. A nail does not evolve; you can’t jumpstart a swarm in your
family, and you probably can’t in a village either. In the natural world, complexity has mathematical signatures
that scientists can look for. Swarms that do most of the heavy lifting online leave behind them a trail of data that
you can search for self-organizing behavior. I am personally involved in an effort to ﬁnd cheap, quick methods to
investigate the matter. If you care about this, we should deﬁnitely speak, there’s not many of us out there.
Internet connectivity and swarm hacking in politics and public policy are not necessarily going to lead to a more open,
accountable and generally agreeable world. Hackers come in two colors: you get the good guys and you get the bad guys. In
2010, Daniel Vaarik and a half-dozen actors made a fake political movement called United Estonia, declaring they would run for
the upcoming Estonian elections. They were worried that ethnic nationalism would be the next big thing in Estonian politics, and
they decided to occupy that position ﬁrst. They had no money at all, but they did have a theater with lots of equipments
(cameras, video editing software, audio stuff...) and a network of suppliers (to print posters etc.). They kept refering to a mythical
leader that would not talk to media, but who would be introduced at the ﬁnal rally that would close the campaign. Almost 10,000
paying people came to Tallinn from all over the country in rented buses to attend the rally. Here is a 4 minutes video that shows
the peak of the event: the introduction of the leader. After the video I will tell you what happened next.
... so we need to stay one step ahead of emergent social dynamics; of the design ﬂaws of our own democratic institutions, and
the mental inertia of the people manning them, i.e. ourselves; and of the bad guys. Failure to do so can result in very, very bad
consequences. Can we do it? I don’t know. I do know that the only way that we can possibly do it is work really hard, be as
rigorous as possible, renounce ﬂuff. It starts small: when you say “innovation”, make sure you have a deﬁnition in mind. When
you say “exponential growth” make sure you know what an exponential curve looks like. Check what you are describing against
that deﬁnition and that curve. Challenge yourselves and others. You will probably end up living in the world I am in: big, scary,
mostly ignores us petty humans. It’s not nice, but in my experience anything else is practically suicidal.