You could have a look at "hierarchical community detection": The general idea is to first detect a small cluster of densely connected (collaborating) vertices - i.e. a Team as you say. In the next step, these clusters are collapsed into one vertex and the algorithm is started again.

http://www.santafe.edu/media/workingpapers/06-10-036.pdf

maybe this way to see a network should helps you or read the book of Melanie Mitchell in the cahpter of networks.

Greg, just out of curiosity, is your data is openly available?

I agree with Greg that the basic approach is straight forward. You simply keep some record of who interacts with whom (according to some operational definition), and then turn that into an adjacency matrix for analysis.

In terms of the theory you are working with, are you familiar with the work of Ron Burt?

Thank you all for your replies. These are all very helpful! I believe you immediately when you say that this approach is pretty standard in many fields. However to my best knowledge it has not been used much to capture team learning behavior. @Karina: thanks so much for the paper, it's really insightful @Greg: I would be very much interested in your operationalization so I would like to read some of it. Which article do you advice in this regard? @David: I'm not (yet) familiar with the work of Ron Burt, but a quick glance at his website shows me that you're probably right, thank you

I have just conducted an 8 month long ethnography for my PhD and used observations plus interviews to map co-worker and friendship/support networks and who hangs out with whom for sociocognitive mapping. I found that my obs differed slightly from interview responses because I noted any interaction with was not work related - but I made note in each observation of what the interaction was about. That way I could select which interactions to map.

Kim. this sounds like a very interesting research method and I would love to read more about it! Could you perhaps share with me your observation scheme? And did you also record the interaction to analyse the quality of the interactions?

You should be aware that there is a rather large literature that people's self-reports of networks do not necessarily correspond to actual observations, where the most prominent authors would be Linton Freeman and A. K. Romney.

"Words, Deeds and Social Structure: a Preliminary Study of the Reliability of Informants." (L. C. Freeman and A. K. Romney). Human Organization, 46, 1987, 330-334.

"Cognitive Structure and Informant Accuracy." (L. C. Freeman , A. K. Romney and S. C. Freeman). American Anthropologist, 89, 1987, 311-325.

Thanks so much for these references! And this is exactly why I want to measure interactions on site instead of using self-report measures.

It has been a while since I read these articles, but I know that there were systematic errors in recall, but these pointed to weaknesses in both observational and self-report data.

Testing my own recall again, I think one of the points from this literature is that people under-report frequent but less meaningful interactions, at the same time that they over-report infrequent but more meaningful interactions.

This goes back to my original remark that any kind of measurement needs an "operational definition" of what you are trying to capture. For example, setting up a video camera might give the best record of actual contacts, but many of those contacts might be quite short and even trivial. Depending on what you actually want to measure, such data could be misleading at best.

If you want to measure certain, specific kinds of interactions, then you need both a systematic way to define when those kinds of interactions do or do not occur (i.e., an operational definition), as well a measurement system that has sufficient reliability and validity. Indeed, the reliability and validity of a measurement system can only be assessed with regard to the definition of what is being measured.

I apologize if I'm doing too much "preaching," but I think it is important to avoid the assumption that observations are inherently superior to self-reports. In particular, data such as importance or value of network ties can be very difficult to assess without self-report.