Hello I will be applying MDS (actually Isomap) to make a psychological "concept map" of the similarities between N concepts. I would like to scale to a large number of concepts, however, the resulting N*(N-1) pairwise similarities is prohibitive for a user survey. I'm thinking of giving people random subsets of the pairwise similarities. Does anyone have recommendations for this situation? My current thoughts are to either 1) use nonmetric/gradient descent MDS which seems to allow missing data, or 2) devise some scheme whereby the data that are ranked in common by several people is used to derive a scaling factor for each person's ratings. Thanks for any advice,
Dear Context Grey, On 15 Jun 2006, at 6:42, context grey wrote:> > I will be applying MDS (actually Isomap) to make a > psychological > "concept map" of the similarities between N concepts. >So actually, how do you do isomap? RSiteSearch gave me one hit of "isomap". I only ask, because I've implemented a working version of isomap (not ready for prime time yet, but a proof that it works). If isomap already is available in R, I won't do anything more with the function. I don't understand the rest of the question, but isomap really may be able to work with NA dissimilarities: just replace them with shortest path distances via non-missing dissimilarities. In fact, you don't need but some ('k') non-missing dissimilarities per item, since that is how isomap works. Your dissimilarity structure may become disconnected, of course, but that's common in isomap. If you mean that your raw data has NA, then you may select a dissimilarity function that can handle NA input and produce finite dissimilarities (I think daisy in the cluster package does this). Somehow I feel I answered to quite a different question than you asked. Sorry.> I would like to scale to a large number of concepts, > however, the > resulting N*(N-1) pairwise similarities is prohibitive > for a user survey. > I'm thinking of giving people random subsets of the > pairwise > similarities. > > Does anyone have recommendations for this situation? > > My current thoughts are to either > > 1) use nonmetric/gradient descent MDS which seems to > allow missing data, or >Not the isoMDS function in MASS. if N(N-1) is a problem, then nonmetric MDS may not be the solution.> 2) devise some scheme whereby the data that are ranked > in common > by several people is used to derive a scaling > factor for each > person's ratings. > > Thanks for any advice, > > _Cheers, Green Power -- Green Power, Oulu, Finland
On Thu, 2006-06-15 at 07:13 +0300, Jari Oksanen wrote:> > > > 1) use nonmetric/gradient descent MDS which seems to > > allow missing data, or > > > Not the isoMDS function in MASS. if N(N-1) is a problem, then nonmetric > MDS may not be the solution.Sorry for the wrong information: isoMDS does handle NA. I remembered old times when I looked at the issue, but isoMDS changed since. Fine work! cheers, jari oksanen -- Jari Oksanen -- Dept Biology, Univ Oulu, 90014 Oulu, Finland email jari.oksanen at oulu.fi, homepage http://cc.oulu.fi/~jarioksa/
Thanks for the isoMDS pointer. I found one implementation of isomap at http://user.cs.tu-berlin.de/~astro/kmethods though the web page suggests the code may be immature. --- Jari Oksanen <jarioksa at sun3.oulu.fi> wrote:> On Thu, 2006-06-15 at 07:13 +0300, Jari Oksanen > wrote: > > > > > > > > 1) use nonmetric/gradient descent MDS which > seems to > > > allow missing data, or > > > > > Not the isoMDS function in MASS. if N(N-1) is a > problem, then nonmetric > > MDS may not be the solution. > > Sorry for the wrong information: isoMDS does handle > NA. I remembered old > times when I looked at the issue, but isoMDS changed > since. Fine work! > > cheers, jari oksanen > -- > Jari Oksanen -- Dept Biology, Univ Oulu, 90014 Oulu, > Finland > email jari.oksanen at oulu.fi, homepage > http://cc.oulu.fi/~jarioksa/ > >