Revealing the weakness of SNA and possibly fixing it, using MAS
A social network model consists of the representation of the target domain in terms of some system of nodes and arcs, plus how inferences about this are going to be made which can be interpreted back to the target. This is a not an analytic results but a contingent theory that can only be validated against independent empical evidence. The approach consists of several stages: (1) the collection of data about the structure and processes in the target; (2) the representation of this in a social network structure; (3) the inference of properties of the network using measures and other results; (4) the interpretation of these inferences back in terms of the target. Properly considered the theory requires all stages and not simply stage (2) (I will call such a SNAT - a social network analysis theory).Given at SNAMAS, part of AISB 2010, Leciester, March 29th 2010.
To validate such a SNAT would require studies to see if there is independent evidence that the outcomes in the target system actually do correspond to the inferences from such a process (as interpreted to the target and given the data collection processes) for the range of targets that correspond to the declared scope of the theory. Unfortunately this is rare, and more frequently a SNAT is only weakly validated against the intuitions of the same researcher that constructed the SNAT. Partly this is due to expense of SNA and independent validation studies, but it also seems to be a result of the way SNA is divided between theoreticians and users. The theoreticians looks at measures and other techniques that can be made of a given network system usually without any reference to observed case-studies - stage (3). The users study observed examples and apply the techniques (frequently wrapped in software to make them more accessible) of the theoreticans to dervice conclusions about what their target systems - stages (1), (2) and (4). Nobody checks that the combined SNAT, all four stages put together, actually works, i.e. subject it to an independent validation.
To demonstrate how SNAT are an inherently difficult and empirical approach, two cases in "Artificial Social Network Analysis" are exhibited. That is where a MAS is studied using SNA methods.
Thus MAS in the form of simulations can be used to probe weaknesses in SNA approaches, showing doubtful assumptions as well as making clear the empirical and contingent nature of SNAT.
- In an apparently simple MAS, where almost all information about the nodes, their behaviour, the social network etc. is known beforehand (everything except the initialisation of the environment), it is proved that there is NO measure that will reliably correspond to the asymptotic importances of the nodes. Given that one can not devise a reliable measure in this ideal and very simple case, this indicates that uses of SNA measures etc. that assume a priori that a given measure is a useful indication of a property of the target system is deeply flawed.
- In a plausible simulation of a P2P file-sharing system, given information that is analagous to what a researcher of social networks "in the wild" would infer, it is evident that the wrong conclusions might well be made. Given this is a simulation it is possible to to check whether the SNAT holds, and despite appearences is found to be lacking. If this is the case for a plausible simulation, how can we take unvalidated SNA analyses of observed systems seriously?
It is suggested that the root problem is the drastic nature of the abstraction step in SNA, from a complex social system to a relatively "thin" mathematical structure - a network. However such abstraction can be staged using MAS simulations. This has the advantage that the chain of reference from model to model is maintained and testable, but at the cost of far more work.