00:00:02
we're here today with Wharton statistics
00:00:03
Professor bashor bataria to talk about
00:00:05
some of his latest research bashor
00:00:07
thanks for being with us today could you
00:00:08
first of all talk to us a little bit
00:00:10
about give us a brief summary of your
00:00:11
research what kind of question you were
00:00:13
trying to answer right so my research
00:00:15
interests are the intersection of
00:00:16
Statistics probability and combinatorics
00:00:19
so recently numerous very interesting
00:00:20
combinatorial and graph theoretic
00:00:22
problems of emerging statistics mainly
00:00:24
because of the ubiquitous presence of
00:00:26
network data and the increasing use of
00:00:28
graph based methods in modern analytics
00:00:30
and as a consequence many interesting
00:00:32
connections have emerged between modern
00:00:34
statistical methods and classical
00:00:36
Concepts in geometry and probability and
00:00:39
you know as a consequence you can use
00:00:40
them to solve interesting problems in
00:00:42
statistics great and tell us a little
00:00:44
bit about in this study what were the
00:00:46
key takeaways that you took away that
00:00:48
you took away from the study the key
00:00:50
takeaways of my research basically the
00:00:52
interplay between computational
00:00:54
complexity which is the time it takes to
00:00:56
to implement one of these methods and
00:00:58
its statistical performance perance
00:01:00
which is how close it is to the
00:01:01
mathematically based procedure it turns
00:01:04
out that many of these graph based
00:01:05
methods are computationally very
00:01:07
efficient so they can be aply used to
00:01:09
solve uh and apply in large data sets
00:01:12
and we have also shown that in many
00:01:14
situations these tests have near Optimal
00:01:16
Performance guarantees which provide the
00:01:19
theoretical justification required for
00:01:20
using these procedures now had graph
00:01:23
based methods is that something that has
00:01:25
not been used as much in the past or had
00:01:27
been looked at with more skepticism mean
00:01:29
why is this
00:01:30
right so so graph based methods have
00:01:32
been used in practice for a long time
00:01:34
but I think what comes out of my
00:01:35
research is the the answer to the
00:01:37
question that why they work so they were
00:01:38
used and they were working fine before
00:01:41
but here is so here we provide some
00:01:43
theory behind why it works giving the
00:01:44
justification of using it in real
00:01:46
problems great and so getting to that
00:01:48
actually like if I'm a business if I'm a
00:01:50
business person or I own a business I
00:01:52
mean how can I apply This research
00:01:55
practically in my life right so one of
00:01:58
the recent projects we're looking at is
00:01:59
what is known as the two sample problem
00:02:01
so imagine I have a situation where I
00:02:03
want to find out whether a set of genes
00:02:06
regulate or affect the occurrence of a
00:02:08
disease so for example to just
00:02:10
illustrate suppose I have uh uh 20 genes
00:02:13
and I have the gene expression level
00:02:15
data from 100 patients who have diabetes
00:02:18
and I have the same expression level
00:02:19
data for 100 patients who are healthy
00:02:22
and the goal is to find out whether
00:02:24
these 20 genes are expressed
00:02:26
differentially so by that I mean that
00:02:28
their expression levels are sign
00:02:29
signicantly different between these two
00:02:31
sets of patients and uh our research
00:02:34
sort of aims to provide the theoretical
00:02:36
understanding and comparison of the
00:02:38
different methods that are deployed to
00:02:41
understand or answer such questions what
00:02:43
are some other applications for This
00:02:44
research right so another interesting
00:02:46
application of our work is in natural
00:02:48
language processing mainly in problems
00:02:50
which try to understand similarity
00:02:52
between words so imagine the word color
00:02:55
which can be spelled in two ways one
00:02:56
with the letter U one without the letter
00:02:58
U and they are the same words but the
00:03:00
words for instance wolf and fox are both
00:03:03
animals but they are very different
00:03:05
words so in this case in spite of the
00:03:08
amount of data we have the support size
00:03:10
which is basically the collection of all
00:03:12
words is far larger than the than the
00:03:14
data set itself so one of our methods is
00:03:17
what we are studying can be used to
00:03:19
analyze such problems as well which I
00:03:21
would think would be pretty interesting
00:03:22
to businesses nowaday just because so
00:03:24
many people are like taking social media
00:03:26
posts and things like that trying to get
00:03:28
data about customers that way right
00:03:30
perfectly yeah yeah yeah sure and so
00:03:32
what's next for This research right so
00:03:34
currently I'm trying to understand or or
00:03:36
analyze the methods for analyzing data
00:03:39
in what is known as the high dimensional
00:03:40
setting where you have say 10,000 genes
00:03:43
and only a few hundred patients and uh
00:03:45
you want to find out something about how
00:03:48
the genes affect the disease or the
00:03:49
patients and for these cases different
00:03:52
new techniques are required and I'm
00:03:53
trying to understand the theoretical uh
00:03:56
the uh background of these results and
00:03:58
how these can be used to find to get new
00:04:00
methods and new algorithms great BR
00:04:02
thanks for being with us today thank you
00:04:03
thank you
00:04:15
[Music]
