Well it may just be you and me Eric. And you can feel free to dropout at any time. Though I assume that Andrew and Trevor are on this list and I know they are somewhat engaged.
This part of the paper explains a few things:
and suppose we believe there to be a ranking of the in- dividuals implied
by the pattern of the unreciprocated friendships so that most such friendships run from lower to higher rank. One possible way to infer that ranking would be simply to ignore any reciprocated friendships and then construct a minimum violations ranking of the remaining network [7, 8]. That is, we find the *ranking of the network nodes that minimizes the number of connec- tions running from higher ranked nodes to lower ranked ones. In practice this approach works quite well: for the networks studied in this paper the minimum viola- tions rankings have an average of 98% of their unrecipro- cated friendships running from lower to higher ranks and only 2% running the other way. By contrast, versions of the same networks in which edge directions have been randomized typically have about 10% of edges running the wrong way. (Statistical errors in either case are 1% or less, so these observations are highly unlikely to be the results of chance.)* The minimum violations ranking, however, misses im- portant network features because it focuses only on un- reciprocated friendships. In most cases there are a sub- stantial number of reciprocated friendships as well, as many as a half of the total, and they contain significant information about network structure and ranking. For example, as we will see, pairs of individuals who report a reciprocated friendship are almost always closely similar in rank.
And they can use the minimum violations ranking to seed their maximum likelihood approach. Without that trick the computational demands would be even more absurd.
So question, does the fact that they can achieve a ranking with 90% of unreciprocated edges directed from lower to higher status people WHEN THE DATA IS RANDOMLY GENERATED worry you... I mean, its good that they can increase this to 98% when the use the ME, but it seems like they have a lot of flexibility. I hypothesize that in schools where people fill out their survey more thoroughly (fewer people with very few edges) that they end up with more edges directed the wrong way. But that the rankings are in some sense truer. Do you think I'm right?
On Fri, Oct 5, 2012 at 2:44 PM, Eric Purdy epurdy@uchicago.edu wrote:
Just out of curiosity, is anybody still reading this thread? What are people thinking of the listhost so far?
On Fri, Oct 5, 2012 at 11:57 AM, Eric Purdy epurdy@uchicago.edu wrote:
Well, we could just say that we'll do what we want, not use their code (except possibly to help us understand the paper), and offer co-author credit to them if we manage to get a paper out of it.
In general, being overly generous with co-authorship doesn't usually have any bad effects?
On Fri, Oct 5, 2012 at 11:34 AM, Michael Bishop michaelbish@gmail.com
wrote:
well its not published yet. i think it was conditionally accepted and probably won't come out until the spring. i'm currently operating
under the
assumption that i can't publish using their ranking without there say
so.
since they said they were planning on doing things similar to what i was planning on doing, and they gave me their code before others have
access to
it (i think), i probably need to consider them coauthors.
but now that i've got the code, its not clear how much more help i need
or
could get from them. therefore i feel like i'm in a grey area where
given
my plans, giving them co-authorship seems generous, but giving them
nothing
more than an acknowledgement seems sorta wrong as well.
On Fri, Oct 5, 2012 at 11:04 AM, Eric Purdy epurdy@uchicago.edu
wrote:
On Fri, Oct 5, 2012 at 10:13 AM, Michael Bishop <michaelbish@gmail.com
wrote:
If I open a second instance of cygwin and run another instance of
this
program, will windows/cygwin know to execute it on a separate core? Brian (the author) just sent me his timings (attached) to give me an idea about what to expect. #3 is the shortest, #50 is the longest.
Probably that would work? I know very little about windows.
Yeah I don't really know what the norms are surrounding this sorta thing. I really don't want to send these guys a huge fuck-you. But it is basically accepted at a real journal... this new one I think? Does its going
to
press change anything?
If it's already published, then we can do whatever we want. And we're not the referees, which is where the real fuck-yous come in. Maybe we should ask them??
Its good to hear their approach seems fairly reasonable to you. It seemed weird to me to bring periodic functions like cosine into it... my intuition is that the density for unreciprocated ties would achieve perhaps 3 local maxima... one somewhere in the middle, and the two "corner
solutions" at
max rank differences.
Yes, the periodicity is a very weird assumption. That's exactly why I was saying polynomials made more sense to me.
I have another friend who suggested eliminating α and β as separate parameters, i.e. don't assume reciprocated ties and unreciprocated
ties
are different, just allow that to emerge... so he suggested a mixture of
3
beta distributions.
If you use polynomials like I suggested, then there's no real difference between alpha and beta being separate or being together. It is possibly helpful at a conceptual level to separate out the parts that are symmetric and anti-symmetric, because they tell you about reciprocated and unreciprocated friendships, respectively.
-- -Eric
-- -Eric
-- -Eric