Complexity and Social Networks Blog

My colleagues and I at the Media Lab recently completed an experiment where we used our sociometric badges to allow people to see how their social network changed over the course of an intensive two week international college student workshop in Japan. The 45 students were placed into teams of 6-8 after the second day of the workshop, where they learned about leadership and recent advances in manufacturing processes and environmental policy. Over the last few days of the workshop the groups competed on a creative engineering task that was judged by area experts. The students wore the badges all day every day for the entire workshop, giving us an unprecedented amount of compete data.

Over the course of the day we would collect data from the badges by transmitting data wirelessly to basestations, and at the end of the day we would download the data from these basestations for processing. Within ten minutes we printed out individual and group feedback sheets for all participants, who then had a reflection session on the feedback and their activies during the day. This feedback consisted not only of a social network diagram for the course of the workshop up to that point, but also of an analysis of the group dynamics patterns that the badges of observed. In our visualizations we showed how much each individual in the group spoke, who spoke after who, and how interactive (vs. lecture-style) each group member was.

There were many interesting things that we learned from this data. Predictably, at first the Japanese students spoke much less than the students from the US, since the workshop was conducted in English. In addition, most of the American students tended to exhibit "cliquish" behavior: American students would only talk with each other. After the first few days, however, these communication problems were solved, and according to qualititative data the feedback was very helpful on this front. At first the assigned group leaders also tended to monopolize conversations, and this was evident in the feedback. They quickly recognized this, however, and soon were engaging other members in meetings and conversations.

The social network diagrams that we provided were also enlightening. Predictably, before the groups formed communication was sparse with few clusters, but once people were placed into groups the intra-group communication amount rose dramatically, although some groups still exhibited sparse communication patterns. Participants and organizers felt this aspect of the feedback to be the most helpful, because it let each group situate itself in the context of the entire workshop and think about how they could cooperate with other group members and other groups more.

Of course, while this is not a formal experiment, the fact that we were able to generate this feedback in real time significantly drove up participation. While participation in the study was a requirement for participation in the workshop, we observed that people were much happier to wear the badges since they could see the effect so quickly. We think that incorporating this type of feedback into studies of this nature is important since it shows the participants that we really are collecting useful data. While it may have changed their behavior, without useful feedback or functionality it seems impractical to expect individuals to religiously wear a sensing devices for weeks or months all the time. We are currently conducting lab experiments to test how effective these feedback mechanisms are at changing group behavior, and we are planning more field experiments to see if we can actually change organizational structure as well.

Posted by Ben Waber at 10:43 AM | Permalink | Comments (0)

I was struck by a presentation by Qiming Lu at the NetSci conference on the dramatic effect of the micro-structure of social networks on consensus formation. Essentially, the results showed that creating random graphs with the same macro properties (clustering coefficient, characteristic path length, etc.) yielded vast differences in the final consensus state of a social network in an opinion spreading simulation.

Lu used social ties between actors to denote influences that individuals have on each other, with actors having certain probabilities of changing their opinion based on their neighbors' states and their current state. In the study, the authors randomly assigned individuals to start with different words for a single concept (called the naming game) to see how many words would exist in the steady state of the system.

In simulations on real world network structures, the authors found that actors would converge to using two words to describe a common concept, while replicating the same macro properties on a random graph yielded a consensus on one word. This has tremendous implications for how we characterize networks, since it points to a lack of a measure to capture certain features of naturally formed networks.

This also leads us to think about how we can combat groupthink, since these results imply that some larger social structures may exhibit some resistance towards groupthink. It is important to isolate these factors so that we can design our organizations and meetings to take advantage of these natural characteristics. Of course we do not randomly start with opinions, but form them over time as a function of those around us. However, these results may strengthen the notion that independent opinion formation followed by social discussion effectively combats groupthink, as has been previously demonstrated in smaller systems.

They have presented some of their results previously, in the paper Dynamics of Naming Games in Random Geometric Networks, but their NetSci paper will hopefully be available online soon.

Posted by Ben Waber at 2:25 PM | Permalink | Comments (0)

I ran a version of the “Bavelas experiments” in my “Building organizational social capital” class last week. The Bavelas experiments were conducted by Alex Bavelas and Harold Leavitt and collaborators in the Small Group Network Laboratory at MIT in the 1950s. The basic experiment is where members of a small group (size N) of subjects are each given a different symbol drawn from a set size N + 1. The task is for the entire group to figure out what symbol is missing. How the group can communicate is dictated by the experimenter, where certain channels are acceptable, and others not. In each round, each subject is allowed to pass a message to one other person. One of the provocative findings from this vein of research was that centralized networks—e.g., where there is a hub that communicates with everyone else but no one else can communicate with each other—tends to perform faster on this task than decentralized networks (in the extreme, where everyone can communicate with everyone else). The basic logic is that in decentralized networks information floats around inefficiently.

Follow up on this vein of research suggests that as the task gets more complex, that decentralized networks actually do better than centralized. An interesting and relevant critique of this research, by Guetzkow and Simon (1955), was that all-channel networks can and do sometimes perform better than hub-spoke networks. That is, the performance of all channel networks was contingent on how they were used. The original Bavelas findings were based on the fact that they were usually used badly.

When I conducted this experiment last year, it pretty much went as expected, except for in one case, where the hub misunderstood the experiment, which resulted (unsurprisingly) in very poor performance by that hub-spoke group. (This highlights that hub-spoke networks are highly dependent on the hub; again unsurprising.)

This year, it went quite differently, where the all-channel network did far better than the hub-spoke network. Both 6 person networks (we did this with two of each and had identical results) performed at their theoretical optimum, which, for the all-channel network was 4 rounds, and the hub-spoke network, 6 rounds. Why the contrast between the years? Because the hub-spoke groups kibitzed about how to communicate before we began the experiment. This only happened because (1) I did not initially explain that no communication (even if it was not about the symbols) was allowed; and (2) the room is so jammed that I could not tell exactly what was going on with the hub-spoke groups, who were in the back (and thus I could not stop them from kibitzing). The all-channel network groups decided, through their kibitzing, that they should all send their information to one person, who would then send the answer back to the group, which would then diffuse efficiently through the all-channel network. That is, per Guetzkow and Simon, they figured out the best way to use their communication capacity.

The interesting lesson to draw here is that all-channel networks (putting the overhead of having more ties aside for a moment) provide a reconfigurable capacity for task-relevant communication. If people are provided (or produce) “programs” about what the task type is, the resulting communication patterns will potentially be more efficient than if the network is hard wired.

It also raises an interesting question, a la Schelling, whether if one allowed some information about the participants in the experiments, and some information about the structure of the network, whether the all-channel network would perform better than it did in the Bavelas experiments. In the Bavelas experiments (as I recall), subjects did not know anything about each other and did not know the configuration of the network. Now imagine, for a moment, that you know what the configuration of the network is, and something about the other subjects; might focal point nodes emerge, to whom most information would be sent, resulting in increased information?

Bavelas, A. Communication patterns in task-oriented groups. J. Acoustical Soc. America 22 (1950), 725-730.

Guetzkow, H., and Simon, H.A. The impact of certain communication nets upon organization and performance in task-oriented groups. Mgmt. Sci. 1 (1955), 233-250.

Leavitt, H.J. Some effects of certain communication patterns on group performance. J. Abnormal and Social Psychol. 46 (1951), 38-50.

Also, see nice summary of research by Steve Borgatti.

Posted by David Lazer at 9:17 AM | Permalink | Comments (4)