References from today's class (collab filtering/LDA/Co-training/jobs etc)

[Subbarao Kambhampati <rao@asu.edu>]

1. The collaborative filtering example I mentioned in the class today
(from a previous exam!)


Don has been married to Mary for 20
years, and they have seen over 300 movies
together. Mary however seems to agree with Don on only 80% of the movies
(in terms of
whether they were good or bad). Of late, Don has been carrying on with
Roxy over at the
office and they have seen some 3 movies together. Roxy seems to agree
with Don a
100%. Now, there is a new movie in town called Gulf Wars: Episode 2
Clone of the
Attack (see 

http://rakaposhi.eas.asu.edu/clone.JPG  for a poster) . He
hasn't yet seen
it, but both Roxy and Mary have and they have differing opinions (Roxy
gives it a 10 it but Mary gives it a 1 
on a scale of 1-10). What does collaborative filtering method say he
should do in this kind of scenario?


[I should admit that in
the exam lot of students rooted for Roxy in their answers.  I have
duly reported them all to Jerry Falwell foundation for
combating moral turpitude.]

2. The LDA/LSI connection

I mentioned that the computing the best directions for LDA wind up being
not as neatly connected to eigen values as LSI/prinicpal component
directions are. To be more accurate, they _are_ connected to eigen value
directions--but not to the eigen values of the variance matrix alone (as
LSI is). Specifically, the LDA directions for a K class correspond to the
top  eigen vectors of the matrix that is given by the product of two
other matrices

inverse(Sw)*Sm

-- where Sw is the intra-cluster variance (scatter) matrix 
and

Sm which is the matrix of covariances of the centroids of the K
classes

For the case when the number of classes is equal to number of data
points, Sw becomes identity, while
Sm becomes the variance matrix of the data--thus LDA naturally becomes
LSI

3. Here is an interesting paper that talks about how co-training compares
to  EM-style self-bootstrapping (the two methods we discussed in the
class). 

http://www.kamalnigam.com/papers/cotrain-CIKM00.pdf (longer version
-- I added it to the readings list)

http://www.kamalnigam.com/papers/cotrain-kddws00.ps (short 2-page
version)

The point they make is that (a) co-training is better than EM approach
(b) co-training works better if you can make the two feature sets
orthogonal
(i.e. conditionally indep given the class).


4. Here is link to Amazon Phx operation with possible job ops coming
up:

 
http://www.a2z-phx.com/careers.php

Here is a link to a google operation of interest too: 

http://www.google.com/jobs/lunar_job.html


Rao