Bernardo Huberman, the director of the Information Dynamics Lab at Hewlett-Packard, gave a talk at Stanford on January 8 entitled, “Information Dynamics in the Networked World.” Huberman covered two somewhat disparate topics: Automatically discovering Communities Of Practice by analyzing email Social Networks, and a general forecasting technique based on markets. (TX)
The first part of his talk was a summary of his recent papers. I’ve alluded to this work here many times, mostly in reference to Josh Tyler and SHOCK. The premise of the work is similar to that posed by David Gilmour in his recent Harvard Business Review article. (TY)
Huberman described a technique for discovering clusters of social networks within an organization by analyzing email. The key innovation is an algorithm for discovering these clusters in linear time. The algorithm, inspired by circuit analysis, treats edges between nodes as resistors. By solving Kirchoff’s equations (which gives “voltage” values for each node), the algorithm determines to which cluster a node belongs. (TZ)
The second part of Huberman’s talk was enthralling: Predicting the near-term future using markets. This is not a new idea. A very topical example of a similar effort is the Iowa Electronic Markets for predicting the outcome of presidential elections. (U0)
The methodology Huberman described (developed by Kay-Yut Chen and, Leslie Fine) aggregates the predictions of a small group of individuals. It works in two stages. The first provides behavioral information about the participants, specifically their risk aversion. Huberman remarked that risk aversion is like a fingerprint; an individual’s level of risk aversion is generally constant. In the second stage, participants make predictions using small amounts of real money. The bets are anonymous. Those predictions are adjusted based on risk aversion levels, then aggregated. (U1)
Huberman’s group set up a toy experiment to test the methodology. There were several marbles in an urn, and each marble was one of ten different colors. Participants were allowed to observe random draws from the urn, and then were asked to bet on the likelihood that a certain color would be drawn. In other words, they were guessing the breakdown of colors in the urn. (U2)
Although some individuals did a good job of figuring out the general shape of the distribution curve, none came close to predicting the actual distribution. However, the aggregated prediction was almost perfect. (U3)
The group then tried the methodology on a real-life scenario: predicting HP’s quarterly revenues. They identified a set of managers who were involved in the official forecast, and asked them to make bets. The official forecast was significantly higher than the actual numbers for that quarter. The aggregated prediction, however, was right on target. Huberman noted that the anonymity of the bets was probably the reason for the discrepancy. (U4)
The discussion afterwards was lively. Not surprisingly, someone asked about the Policy Analysis Market, the much maligned (and subsequently axed) brainchild of John Poindexter’s Information Awareness Office. Huberman responded that the proposal was flawed, not surprisingly suggesting that this technique would have been the right way to implement the market. It was also poorly framed and, because of the vagaries of politics, will most likely never be considered in any form for the foreseeable future. (U5)
I see many applications for this technique, and I wonder whether the Institute for the Future and similar organizations have explored it. (U6)
As an aside, I love how multidisciplinary the work at HP’s Information Discovery Lab is. I’ve been following the lab for about a year now, and am consistently impressed by the quality and scope of the research there. (U7)