Complexity and Social Networks Blog

First, consider a less problematic example - email logs. Here, the network is formed by the collection of asymmetric sender-recipient ties, and no content is coded. Although "flame wars" are a well-known phenomenon of electronic communication, it remains unlikely that the pattern of one's email network for friends is similar to that of one's enemies or even acquaintances. Indeed, VisiblePath's goal is to infer trust relationships from email, IM, and scheduling logs. That is, information about the *content* (or consequences) of ties can be inferred from network structure. This analysis provides novel insights that can be difficult to obtain by other means.

Consider another example, the person-gadget bipartite network of professional reviewers of tech gadgets harvested from the text analysis of their reviews. Absent the content of the review, a network representation of the presence of review ties is unlikely to provide any new insights. The domain of interest or expertise of a reviewer and his/her community of peers could likely be inferred from the network, but these could be found at least as easily by examining the staff rosters across the periodicals used as sources. Absent information on the content of the tie, where the content varies independently from structure, novel insights from exclusively structural
analyses will be more limited.

Now take the example of automated data harvesting of blog entries, for example, where interpersonal networks could be constructed based on author-subject ties (let's limit the subjects in this example to other individuals identified by name in the text of the blog entry). Entries can be about topics an author finds inspiring or infuriating. Their subjects could be just as likely to be praised as pilloried (especially on political blogs). In such a case, then the range of possible inferences about the consequences of these ties is constrained as in the previous example.

In these problematic cases, it is decidedly NOT necessarily true that the network does not matter. Rather, the problem is that divorced of the content of the tie, it can be hard to know whether or how the network matters. Null effects from network analyses could arise from the lack of any effect, or they could indicate that the network definition was insufficiently sensitive to the content of the ties. The solution is to encode some of the content within the network. So for the examples above, including the number of stars as a tie weight for reviewer-gadget ties, or the ratio of positive to negative adjectives in the text around a subject's name in the author-subject ties of blog entries, could provide for more novel insights from a network analysis.

As more network analysis is conducted on data harvested in automated ways, it is important to consider the possible importance of tie content, and how it may be incorporated into structural representations to best support revelatory analyses.