Honest Signals Online

So I’ve been reading Sandy Pentland’s recent book, Honest Signals: How they are Shaping Our World, and got to thinking about how honest signals are re-mediated in online environments. For the sake of a quick summary, Pentland’s group measures four honest signals automatically using what they call a sociometer, a device worn around the neck designed to record audio, accelerometer, and proximity data for a person. The four honest signals they detect are influence, mimicry, activity, and consistency. Using these signals they can predict all sorts of behaviors such as bluffing in poker, exchange of numbers in speed dating, and success in salary negotiation.

The model Pentland posits suggests that social circuits can be used with idea markets to make highly complex group decisions by aggregating unconciously held beliefs about the risk / reward tradeoffs of alternatives. Bets are placed on various ideas through a show of interest signalled via activity level. The alternative with the highest overall reward for the group with the least risk is selected by integrating the information from the group members.

One problem I see with this model is that it doesn’t account for partial information. It’s feasible that if group members do not have comprehensive information (or have experienced enough similar situations to be intuitive about it) then their signalling about interest may be misinformed. They may not have the “big picture” which I think is important in decision making. The argument could be made that the big picture is integrated from each individual’s slice of the experience / understanding, however this misses contextual factors between parts of a complex problem. For instance, if proposition A interacts with proposition B in complex ways, then the context of proposition B (i.e. the state of proposition A) is important for making a decision about proposition B. This suggests to me that there are problems for which the group may not have enough aggregate experience to make an objectively good decision. The decision will be good with respect to the risk / reward perceptions of the individuals but could still harbor major risks or diminished rewards if perceptions had been for instance manipulated externally.

Caveats to the decision mechanism aside, I think Honest Signals provides interesting food for thought on how social signals are transmitted in online mediated environments. All of the studies presented by Pentland considered only meat-space social interactions, but online interactions are a major source of socialization now including everything from online discussion forums to World of Warcraft. So how are influence, activity, mimcry, and consistency remediated? And do they mean the same thing online?

In online discourse we’re clearly missing things like physical activity signals, gestures, vocal intonation and prosidy, and posture, but perhaps there are analogue signals that can be detected as proxies.¬† Further confounds in analyzing online discourse are that its features must be extracted from written language rather than from lower level signals and that written language is concious (and could be manipulated) whereas other honest signals are unconcious.

First of all, we need to clarify whether we’re talking about synchronous versus asynchronous communication online. For the case of synchronous communication (e.g. instant messaging), there seem to be many analogues that could be measured in the discourse. Activity could be the amount of typing done in proportion to the whole, influence could be a measure of the turn-taking behavior and likelihood to induce a response in the interlocutor, mimicry could be a measure of similar types of language used (e.g. leet, emoticons etc.), and consistency could be measured via the prosody of typed exchanges. These various features for detecting signaling in synchronous communication could likewise be used to predict social roles or even the success of IM mediated customer service “calls.”

The case of asynchronous communication seems more troublesome in terms of translating activity, influence, mimicry, and consistency. Activity level could still be measured as a response rate, or consumption rate (how many posts you read). Likewise in something like a threaded discussion forum, influence could still be measured by looking at discourse structure and the tendency for some people to seed deep and wide threads of discussion. Similar to the synchronous case, mimicry could be measured by the amount of linguistically similar constructs used in thread exchanged. Finally, consistency could be problematic since in non-realtime communication it’s unclear whether prosody is meaningful or interesting. At the same time, time-stamps in asynchronous communication could still indicate a pulse for consistency.