GroupLens Research

I finally got around to carefully reading "A Theory of the Critical
Mass..." by Oliver, Marwell, and Teixeira. Now I'm asking: what took
me so long?

The article formalizes the notion of critical mass in collective
action. It identifies two main independent variables that can
influence the "probability, extent, and effectiveness of group actions
in pursuit of collective goods":

• The form of the "production function" that relates "contributions of
  resources to the level of the collective good". Two important
  categories of production functions are: (a) decelerating: the
  "first few units of resources contributed have the biggest effect on the
  collective good, and subsequent contributions progressively less"; (b)
  accelerating: "successive contributions generate progressively
  larger payoffs; therefore, each contribution makes the next one more
  likely."
• The "heterogeneity of interests and resources" in the population of
  potentially interested actors.

The authors then show that the problems and opportunities for
collective action are very different for accelerating vs. decelerating
production functions and for homogeneous vs. heterogeneous populations
of actions. I'm not going to summarize the findings: the paper is a
joy to read, so I mostly want to urge you to do that.

However, there were a couple of ideas that I found particularly
relevant to issues in open content systems that I care about, so I did
want to mention them.

First, this work looks at critical mass in "public" goods, where all
the value is created by a group of people. This is true for many open
content systems: Wikipedia and OpenStreetMap are two good
examples. However, this isn't true of other systems, including our
Cyclopath bicycle routing system. Cyclopath began with a nearly
complete transportation map created from Mn/DOT data and with a good
objective route-finding algorithm that did not require user
input. While we have shown that user input improves route-finding
significantly and that algorithms based on user input are better than
purely objective algorithms, I think it's fair to say that most of the
value of the Cyclopath "good" already was present before any user
contributions were made. It's interesting to consider how the concepts
of this paper can be applied to a system like Cyclopath.

Second, Oliver at al. show that with decelerating production
functions, the optimal outcome would be achieved if the *least*
interested people contribute first and the *most* interested people
contribute later. This obviously isn't the way it usually works. They
point out that one way to make this happen is for the most interested
parties to "hold back"; perhaps they can offer "matching
contributions" to entice less interested parties to contribute early
in the process. This might suggest new strategies for
intelligent-task-routing-like strategies to elicit participation in
open content communities.

Third, many of the illustrative examples the authors give concern the
different opportunities for collective action in "upper middle class"
vs. "lower income" neighborhoods. I wonder: what's the equivalent of
an "upper middle class" open content system?

Fourth, the notion of "interest" presumed here is one of direct
tangible personal benefit: if I give N dollars, I'm increasing the
chances that I'll receive M dollars (M >> N). However, we know that
many contributors to open content systems (and many 'volunteers', too)
contribute for other types of reasons, e.g., they "believe" in the
public good, they are altruistic, or they want to build a
reputation. For example, in Cyclopath, our most active editors don't
request many routes. For another example, other researchers have shown
that there are many users in discussion forums who just answer
questions and don't ask any of their own.

Fifth, finally, and simply, I'd like to empirically measure the
production function in various open content systems. I suspect that in
many cases it is decelerating: i.e., early units of contribution are
proportionally more valuable. I'd also like to measure this for
individual users. Doing this calculation requires a way to measure the
global quality of an open content system as well as the quality for a
particular user. We can do both of these for Cyclopath. We can do the
latter for MovieLens... not sure about the former.