Wednesday, July 23, 2014

What I learned about this week - Social Physics

This week I read Social Physics by Alex Pentland, who runs MIT's media lab. His research has to do with understanding the way ideas move/interact with social networks, sort of like how normal physics is about understanding the way that particles move/interact in physical setups. The book talks about the basics of social physics, and how we can use it inform the way we act in groups, organizations, and societies. Overall, it was a little bit cursory and I'm not convinced of all of its claims, but it touched on several provocative ideas. It's also been a while since I have written a blog post, so I thought I'd type up some of my notes.

Social Physics

p. 26 - The most creative people are the ones who draw many ideas from diverse sources, and who then bounce these ideas off of the other people they interact with. The process - which we call "idea flow" - consists of exploration to find ideas and engagement with peers to sort through them.

p. 37 - People in social networks form many of their ideas through exchanges with other people. This can sometimes lead to many people having highly correlated beliefs, without realizing that they didn't form their opinions independently. We can think of overconfidence/groupthink as results of unawareness of these sorts of feedback loops.

p. 38 - We can provide financial incentives for certain people to exchange ideas outside of their normal social networks, so as to fine tune increase the informational content of idea flow. This was done in an experiment with stock traders, and it lead to everyone's profits increasing significantly. It might also help to avoid the feedback loop problems.

p. 67 - Social network incentives can leverage peer pressure by giving rewards to someone's friends (and nothing to the actor!) in response to some desired action. In some contexts this works much better than just paying the actor.

p. 71 - We respond more to information about what our friends do than what people we don't know (or large-group averages) do, sometimes to the extreme.


p. 88 - In an instance of modeling data on group decision making, the quality of idea flow (roughly, the diversity of the members of conversations within the group) was as important a factor in predicting the group's collective success as was the skills of the groups members.

p. 114-5 - When we obtain information from groups, it's important that we don't double-count dependent signals resulting from idea-flow within the group. Three methods for trying to only count independent signals:
  1. Ask people to tell you not only their beliefs, but also what they think everyone else's beliefs are. Use this to isolate the independent information.
  2. Look for the people who are best at predicting everyone else's beliefs, but who also have different beliefs from these; presumably they have the most data. Pay more attention to them.
  3. Model social networks and the idea flow and influence through them; then infer.  (This is more intensive, and also is supposed to work better.)
p. 116 - Group members with high social intelligence can improve group performance by encouraging more balanced discussion / flow of ideas. We can enhance individuals' social intelligence by showing them graphics about how much each group member is communicating with each other.

p. 124 - In the "Red Balloon Challenge,"DARPA offered a prize for whichever group most quickly could locate ten red weather balloons placed in random-ish locations in the US. The winning group used a social network incentive technique, where they build a huge network organization by promising fractions of the prize money not just to members who found balloons, but also to the members of the organization who had invited those members, and also some to the members who had invited the inviters, and so forth. The moral of the story is that by using network incentives, we can not only accomplish tasks, but also mobilize robust task-solving organizations.

p. 129 - The key to social network incentives is that people are not just motivated by financial gain, but also by social pressures whose strength is proportional to the strength/trust of the network connection.

Data-driven cities

p. 138 - The path to a "sustainable future" could involve using technology to build a "nervous system" that can tell what's going on in the environment (e.g. by inferring from google searches), making models of societies' demands and reactions, and then acting accordingly. Getting this data will be pretty easy. The main challenges are (a) using big data to build successful models of society and (b) passing a "New Deal on Data" - a new systemic and legal framework to ensure privacy, stability, and efficient government.

p. 143-7 - Some examples: 
  1. Model traffic patterns to change stoplights, etc. based on live data (from smartphones) about where people are. Improve flow, reduce accidents.
  2. Track disease outbreaks in a city by observing changes in web behavior. Use GPS data to incorporate with traffic models and reduce the city's exposure to the disease.
p. 162-4 - More exploration of different stores is associated with increased population, greater diversity of available goods, and higher GDP. When people have abundant resources, they don't just explore in order to discover more efficient behaviors; they actually have a preference for exploration and curiosity (rich people in New York City like to try out new restaurants in new neighborhoods, even those they don't expect to be better than their favorites). This is a great characteristic, insofar as exploration is key to idea flow and the related innovation.

p. 166 - Population density, idea flow, and GDP are highly correlated in cities. Some people think there is a very clear causal relationship: more density -> more interactions with new ideas -> better discoveries and innovations -> increased GDP.

p. 185 - In sciences like physics, it's comparatively easy to isolate variables and therefore do causal experiments. This is not possible for data collected about entire cities - but there often is not enough information in a data set to tell us which of many possible explanations is correct. The solution: "living labs" in which participants - who agree to have large amounts of data collected about their day-to-day lives - can participate in experiments that allow for the testing of causal relationships in complicated, interactive settings.

A new approach to economics

p. 197 - The idea of a market - where everyone has the same access and ability to compete - is a gross simplification of the real world. Really, participants have different social connections, information, ability to trade, etc. - we really participate in "exchange networks" that function through our relationships with other people. (The idea of distinct classes, defined by market participation, is similarly outdated; we are members of interwoven social networks.)

p. 198 - Exchange networks are worse than traditional markets when they link participants to such an extent that failures cascade through the system. However, they can lead to more stable and trusted buyer-seller relationships.

p. 201 - In some game theoretic modeling of exchange networks based on fewer, stronger relationships between market participants, it is found that the market is more stable and distributes surplus more evenly between participants.

My favorite concepts from the book

  1. Thinking about idea exchange in terms of networks and flow, and the implications for dependence of beliefs and overconfidence.
  2. Social network incentives leverage the fact that people don't just respond to money; they also respond to social pressure.
  3. In theory, we could use smartphones, etc. to collect very fine-grained data about people, and use this to update policies in real time.

No comments:

Post a Comment