A computer model that detects who is influencing whom in a group discussion, can accurately predict who is likely to speak next
One fascinating question that occupies social scientists concerns groups discussions. The problem is to determine the nature of the interaction between individuals and in particular, who influences whom.
Various breakthroughs in network theory and agent-based modelling have revolutionised researchers’ understanding of these processes. One approach for analysing online discussions is to look for the set of keywords that define a topic of discussion, record the various instances in which these words appear and then study the links between the sites that use them: which came first, who links to whom and so on. This data can then be used to construct a network of influence.
While this has been hugely useful, it’s hard to escape the sense that it fails to capture the true dynamics of influence, the way the balance of power and influence within a group shifts from moment to moment as a discussion evolves. If you’ve ever participated in a face to face group discussion, you’ll know what I mean. (And if you haven’t, where have you been?)
Today, Wei Pan and pals at the Massachusetts Institute of Technology in Cambridge take an important step towards righting this situation. They’ve built a reasonable model that simulates the ebb and flow of influence between individuals during a group discussion. They do this by creating a conventional model of the network of influences between individuals and then take into account that these influences change in time.
What’s impressive about this approach is that it has predictive power in the real world. Pan and co applied the model to data taken from real world discussions in which groups of four people took part in brainstorming and problem solving sessions, either face to face or in separate rooms.
The question Pan and co try to answer at each point in these discussions is: who is going to speak next. Humans listening to these discussions get this right about half the time. Presumably, they are able to use various cues such as the topic of conversation and the inferred emotional state of each speaker.
Pan and co’s algorithm does significantly better than this, correctly predicting the next speaker between 55 and 67 per cent of the time. And get this: it does it using nothing but the volume of speech to determine the patterns of influence between individuals.
That’s impressive but the team has even more ambitious plans. Groups of four are relatively simple. But what of groups of hundreds or thousands? “Our immediate next step is also to apply our approach to larger and longer individual human behavioral datasets which we are currently collecting,” they say.
Beyond that is the question of how this data and the model they’ve developed can provide feedback in real time that improves the performance of the groups. Is it conceivable that a system like this could coach individuals involved in discussions in a way that improves the outcome? And if so, how long before we see the iPhone app?
Ref: http://arxiv.org/abs/1009.0240: Modeling Dynamical Influence in Human Interaction Patterns