How to think in networks

We use the word “network” so regularly in our vocabulary, from a list of the people we know to a type of cable, that it can be a difficult word to define. We define a network as a group or system of interconnected people or things.

While we tend to think of networks as online creations, networks can be an automotive supply chain, our telephone system, or a organized group of people. If it seems broad, it is important to point out that the vast majority of what we see in this world is not a network. A chair, an unconnected piece of software, a collection of rocks, a random collection of people are just a few of the examples that are not networks.

Networks have three characteristics:

1. A network is a collection of elements and connections. It is not a single entity or a series of disconnected entities.

2. A network is greater than the sum of its parts. It is not a pile of items that, when grouped together, have no new characteristics (such as a handful of sand).

3. A network has a function or purpose. While this may not be as obvious as the other first two principles, our definition requires a network to have some reason for being. A network might carry electricity, help friends talk, or bring together buyers and sellers.

In 2016, the Prophet Brand Relevance Index (BRI) surveyed 15,000 people to gauge which of 300+ brands in 27 industries they could not live without. Network companies (Apple, Amazon, Google/Android and Netflix) took the top 4 spots. Why do consumers increasingly prefer network-operated services to non-network operated services?

Online networks save you time. A 2018 study showed that ride share services like Uber and Lyft can save users 15 minutes vs hailing a taxi. An additional study showed that 90% of rideshare trips were faster than taking a train or bus, with some differences based on city geography. Similarly, a 2011 study showed that look for a job on an online jobs network reduces the time spent unemployed by 25 percent. Of course, we don’t need data to appreciate this convenience.

Networks have even made more complex tasks, such as meeting your spouse, more convenient. Anyone who has booked a hotel room, sent a running late text to their spouse, got lost and turned on GPS, or need to get a quick answer on a movie time has seen how an online network can make complex tasks seems trivial.

For business, companies have adopted networks have seen large gains in productivity, or the output generated for the same amount of people. For example, one study found a doubling in productivity in trucking and logistics companies due to GPS adoption, and other studies have shown impressive results in recruiting, learning, marketing and customer support . Finally, even in government services, networks have led to important time savings in critical moments due to networks. The first question 911 operators ask is for your location; mobile phone networks now automatically provide this and save valuable seconds. Similarly, the AMBER Alert system was set up by the government to find lost children; mobile networks now communicate this information to all phones via push notification, with 45 children rescued via the wireless emergency alerts.

Not only do digital networks save you time, they also save you money.

One way to appreciate the economic advantages of online networks is to look at how networks have impacted the costs of other industries that have adopted them.

• Online networks have made it cheaper to move money. For example, rates for wiring money from Mexico to the US dropped from 15% to 6%. The impact? From 2000 to 2009, the total value traded across borders rose from $6.5 trillion to $15.5 trillion. The percent of foreign direct investment rose from 6.5% of the world’s global economy in 1980 to 30% in 2010.

• The cost of air travel has fallen dramatically in the last 30 years - in a large part driven by better online logistics systems. The impact? Today, over 320 million people fly to attend professional events and meetings alone, and the number is growing. Even more important, the UN estimated a 37% increase in migrant relocation in the last 20 years, with over 200 million migrants now living across the globe.

• The cost of moving cargo on transportation networks is 10 times lower than 1950. The impact? In 1990, the world’s total exports and imports comprised 39% of the global economy. By 2010, it was 56%.

Why are networks so good at reducing costs? Networks have three distinct advantages.

Operating Costs

Networks is dramatically reduce operating costs by eliminating the need for multiple top-down processes and structures.

Retail marketplaces serve as a fantastic example of how online networks can reduce costs by flattening operating models. Not too long ago, we went to a local store because it was easier for the store to aggregate milk, vegetables, shaving cream, etc. from various suppliers than for anyone to do so individually. That local store was a network that connected buyers and sellers. Then Walmart moved to town. Walmart used smart logistics software to scale the coordination of goods across the nation. In doing so, they reduced the costs of having local planners and managers coordinate goods A few years later, Amazon launched. Amazon’s software could match buyers to sellers online – eliminating the need for planning inventory, customer service personnel and floor space. Today, Amazon manages a much more efficient operation than Walmart. Walmart employs 2.2 million people to achieve $482 billion in sales, or approximately $219,000 per employee. Amazon employs 150,000 people to achieve $89 billion in sales, or approximately $593,000 per employee.

Uber and Airbnb serve as more recent examples of networks changing industries via operational efficiencies. Amazingly, the largest taxi company owns no taxis or dispatchers. Why would an organization need a dispatch system and vehicle inventory when Uber can coordinate drivers and riders directly? Similarly, the largest hotel chain owns now buildings. Why would an organization need rooms, when Airbnb can directly coordinate space and renters?

The power of networks to reduce operational overhead and still effectively reach more people has appeared outside of corporations too. In militaries, the ability to communicate across smaller groups has made it easier to be a guerilla army. According to a Harvard study, the weaker side of a asymmetric conflict achieved its goals 12% of the time between 1800-1849; between 1950-1998, for a remarkable 55% of the time, the weaker side prevailed. While multiple factors contributed to this change, it is probable that the ability to organize without having centralized units was one of them. In finance, hedge funds could quicker adjust to market trends in volatile markets than large banks, in a connected financial world. In the second half of 2010, highlighted by the US economic downturn, the top 10 hedge funds out earned the top 6 banks. Even in religion, Pentecostal churches, small churches that can cater to local markets, have out-maneuvered organized religions built on centralized approval and planning. In Brazil, Pentecostals made up 49% of churches in 200 vs. 5% in 2006.

Management Costs

While networks can eliminate layers of management, it cannot fully eliminate the need for managers and decision makers. For those remaining, networks can make managers most effective.

Networks provide feedback loops that can automate key decisions for managers, reducing the amount of managerial tasks required. Ebay’s managers do not have set prices, an algorithm automatically suggests prices. Ubers managers do not have to incentivize and motivate new drivers to hit the road, an algorithm automatically analyses data and an app gamifies mileage. Marketing manager who buy advertisements on Facebook or Google have their budgets allocated for them.

For those decisions not automated, networks provide the data to allow managers to better measure and monitor operations. Not only do networks provide an immense amount of data, they have created tools to visualize and make sense of that data, allowing managers to test various hypotheses and model systematic outcomes.

Access Costs

Often, the most expensive part in distributing a service is the marketing required to reach community members. Online networks can drastically reduce the cost of reaching people via network effects.

In the industrial age, supply economics (such as the ability to buy parts at cheaper cost with volume) allowed some companies to grow quickly. In the online age, some networks experience a unique phenomenon, aptly called the network effect, that allows them to grow quickly and last a very long time.

When a network effect is present, the value of a product or service increases when more people use it. In same-side network effects, the more of the same kind of people who join the service, the stronger the network gets. For example, the telephone network became more valuable with every home that got a telephone. Soon it became very difficult to operate without a telephone because every household had one, creating a tipping effect where talking using other technology would be pointless. In cross side network effects, the network gets stronger when members on the other side of the market increase. For example, the more fans who joined Instagram, the more the service became valuable to photographers – even if other tools had better photo editing software.

Once a network effect takes place, the cost of reaching people fall dramatically, as messages can be targeted to specific groups of people. Facebook for example, allows people to uniquely target content on traits including, gender, relationship stats, education, job titles, places of work, current location, radius around your store, hobbies, interests, activity on the site, entertainment preferences and more. The site can also track you through other sites, as long as browser cookies are enabled and you remain logged in on that device. On private networks, how much data is collected and what users know leads to many questions, but it is clear that this data does bring down the cost to reach customers. Advertisers increasingly see targeted online ad spend as the most cost effective method to acquire customers. Global online advertising spend grew from $0 in 1995 to over $200B in 2017, surpassing television ad spend and growing much faster than any other type of advertising, with Facebook and Google owning roughly 85% of all spend.

Online networks can offer new ways of creating trust, increasingly important to customers in an era where trust in many institutions has never been lower.

When Edelman, a public relations firm, asked participants across 20 countries who they trust, a key themes emerged: “A person like me.” Networks have an inherent trust advantage over non-networked services in their ability to facilitate interactions with people you inherently trust.

How do networks facilitate peer trust? First they make it easy to access people you already trust. Research shows that most of us talk to 7-15 people, with 80% of that time being spent with 5-10 people. These “strong ties” have a huge influence on us, impacting whom we vote for and what products to purchase. For example, 82% of Americans consult online reviews when buying a new product., with 40% saying nearly always turn to reviews before a new purchase. As these little interactions add up, as a system, networks often become the trusted place we all turn to discuss which problems to solve and which solutions are preferred to address those problems. As one example of little interactions leading to larger change, in March, 2013, 3M people changed their Facebook profile to the red equals sign to support marriage equality. While this might seem like a small action, researchers found that the likelihood to change one’s profile (a proxy for also showing your support) was greater with more exposures to changed by friends.

Networks don’t just increase trust by enabling us to interact with people we know, they also increase trust by letting us build a trusted reputation and interact with people we don’t know. Ratings and reviews are key mechanisms used by networks to crowd-source reputation. We are 3x more likely to book a hotel with a 5 star rating than a 3 star rating. We are much more likely to purchase a product with a 4 star rating than a 3 star rating (4 star ratings get 11.6x more orders than a 3 star rating). We are even more likely to let someone outside our comfort zone stay at our house when his or her ratings are good; a Stanford and Airbnb study found that “having enough positive reviews can help to counteract homophily, meaning in effect that high reputation can overcome high similarity.”

At first, the idea that we trust networks might seem very wrong. After all, we all have seen things that are clearly untrue on Twitter, Facebook or any other network. The facts back this up. A 2018 Pew study found that only 3% of Americans now say they have a lot of trust in the information found on social networks, and the majority of Americans do not trust that their information is being kept safely. And 74% of people did not realize the extent of the data that Facebook collects.

While we may not trust the institutions that run networks, we still trust the individuals we know that make up the network. This is because networks have created new “horizontal system” of trust, such as allowing you to keep track of people you have met, the close connections of people you trust and rating systems voted up by the community. Horizontal systems, where trusted information can spread between peers (vs “vertical system of trust”, where trusted information is spread by institutions and experts) is a massive advantage networks have in solving communal problems. Indeed it is when networks fail to manage their systems of horizontal trust - for example, allowing bots to create fake ratings and reviews, allowing too many connections per person or attempting to add vertical verified systems - the network becomes untrusted.

A network is composed of a set of elements called nodes. These nodes can be either homogeneous, where all nodes are of the same type (such as Skype users where everyone has a number) or heterogeneous, where the nodes are different (such as buyers vs. retailers on Amazon or Ebay).

Edges connect nodes and are the other element required for a network. Edges can be unidirectional (such as Twitter, where one person follows another) or bidirectional (such as Facebook, where two people friend each other)

Small tweaks to how we represent edges can bring about very different networks. For example, on any given road, we can add labels making the road one-way or two-way. We can adjust parameters, adding tolls that change traffic. We can change the information flow rules, for example adding traffic lanes. We can add behavior feedback, such as signs that measure a car’s speed against the limit. We can add buffers and delays, such as on ramps that help us keep traffic flowing. We can add feedback loops, such as billboards that tell us dynamic toll prices based on the day’s traffic.

These seemingly minor network changes can lead relieve hours of traffic and reduce thousands of accidents.

One way to describe a network holistically is by size. Typically, we define size in number of nodes (i.e., buildings) and edges (roads). For instance, social networks can be described by their number of users – to show Facebook is bigger than Twitter. By increasing the size of the network (growth), you can change the network.

Density refers to the ratio of the number of edges to nodes. Some networks are dense, like the city grid that has each block connected. Some networks are not dense, like a rural highway that connects only a few destinations.

The clustering coefficient is a measure of an "all-my-friends-know-each-other" property. A high clustering coefficient for a network means that it is a tightly knit community.

A network can be organized in multiple ways. These include tree and hierarchical structures (typically man-made, such as organizational charts), random structures (in which each node has on average the same amount of edges, such as city grids), “small world” networks (with high clustering, such as social networks) and scale-free networks (where there are some nodes with a lot of connections and many more nodes with few connections, such as websites).

In some networks everyone likes each other. In others, two groups of mutual friends hate each other. Amazingly, no other configuration of balance exists. Known as the Balance Theorem, Frank Harary realized in 1953 that either all pairs of nodes are friends, or “else the nodes can be divided into two groups, X and Y, such that every pair of nodes in X like each other, every pair of nodes in Y like each other, and everyone in X is the enemy of everyone in Y.”

Sometimes two networks can interact (such as a social network sending traffic to a messaging network). When two unique networks increase the usage of each other, they are considered complementary

Sometime we make a change to the network and the true impact of the change takes a while to play out. For example, removing a bus stop may make busses do a route faster…until ridership drops as those who needed the stop find new transportation options. Not understanding or accounting for time delays are one cause of bad decisions

One common pitfall of a network is that people sub-optimize for a part vs a whole.

Sometimes this is unintentional. People simply don’t realize that a subsystem is creating negative tradeoffs, either due to how the system is measured, focus or rose colored glasses.

Sometimes this is intentional. A manager wants to grow their P&L or influence. They end up making decisions to grow their area at the determent of a network.

How to optimize for the whole remains the hardest leadership challenge in the age of networks.

The worst thing that can go wrong in a network is a negative feedback loop. Here a change is magnified to create a much worse outcome for the system than possible.

In a power plant, a change to one variable might lead to the system overheating and combusting. Lowering gas prices might lead to less fuel efficient cars and more pollution. A switch in Ohio breaks, leading to the blackout of the east coast (2003).

Actively managing against negative feedback looks (breakers, easy rollback of changes, etc) can be seen as overhead…until you have had to deal with a negative feedback loop.

Empty calories occur when you think you are improving the network, only to actually create zero change. Like sugar, you are gaining weight but not getting nourishment (increased org resources not helping a network actually get better)

Of all the pitfalls I have seen, this happens the most. People will spend time on something they wish worked, may in theory work, but doesn’t. Examples include promoting changes with zero retention in a social network (ie does not lead to more value created) or financial changes that increase stock price with out creating something new (ie does not lead to more value created).

Outdated models are dangerous because they present the wrong view of the world. This leads to bad decisions

No model is perfect, but in a complex network a model is essential to best predict cause and effect. A bad model can be worse than no model because it leads to false confidence. For example, if your model makes you believe that a traffic light will reduce traffic - lights will be placed in the wrong place; larger capital expenditures, traffic cannot be improved. Thus, you’d lose money and made traffic worse vs not making any changes.

The most annoying problem to solve in a network is conflicting goals.

This is a common problem You want a social network to be trusted and engaging, a marketplace to be efficient and profitable, air travel to be affordable and enjoyable. The tradeoffs are everyone

Principles and reduced goals can help, but the biggest way to manage a network is to accept that conflicts will occur. An escalation process and understanding that you will never have as much simplicity in decision making as you’d like are essential to working on networks

A network can be understood by systematically looking at the nodes to see differences in node influence and centrality. For example, marketers may want to see if a social network is heavily influenced by a few celebrities or via close connections.

A network can be understood by systematically looking at edges to find the longest, shortest and average path. For our telephone system, for instance, this can help us find out how vulnerable the network may be.

A/B testing is a method of comparing two different treatments and measuring results. For instance, suppose we wanted to increase the open rate of an email; two emails can be sent – with some percentage of the population seeing one subject line and another seeing another subject line. We can then know which subject performed better. In a similar way, suppose we want fewer people to use electricity, we can test various billing design and pricing algorithms to see which one is most effective.

Sometimes changes are too big to measure in a simple a/b test. We launch a new product or target a new segment. Here we can look at cohort analysis and measure, over time, how a group’s behavior compare to the whole over a lifespan. For example, if a national anti-texting campaign is launched in March, 2014 and aimed at new drivers when they get their license, we can measure that groups’ road accident vs. the rest of the population to see the impact.

Funnel analyses are more a targeted way to measure conversions that can be used with other forms of analysis. Often, the goal is drive a direct behavior – be it converting more people to paid subscribers on a retail network or getting more people to sign up for healthcare on a government website. Funnel analysis looks at each step from start to transaction and measures where people are not dropping off.

Sometimes we need to know how a network might react to a change before launching it live. For instance, it is difficult to test privacy-based changes on a social network for a small population, as it breaks the members’ trust. To predict changes, we can create a set of equations based on data from previous experiments. For instance, we can estimate that similar privacy change led to a 0.3% drop in page views or a 2.1% increase in vehicle traffic across the network. As this change impacts only 50% of users, we can estimate half the impact to the network.

Some systems are too complex to model with equation. As opposed to Equation Based Models, Agent-Based Models allow us to model the actors in the system (people, cars, etc) and see how they would react in a system. By encompassing simple rules in agents, we often find non-linear behaviors. For instance, it is difficult to find the equations that lead to how birds fly in formations, but a few simple rules programmed into the agents representing birds lead to the same flock behavior.