I’d like to talk to you today about Big Data
Data
Data it’s a term we’ve probably all heard a
it’s a term we’ve probably all heard a
it’s a term we’ve probably all heard a lot recently one of the things that’s
lot recently one of the things that’s
lot recently one of the things that’s really interesting about it though is
really interesting about it though is
really interesting about it though is that it seems to mean very different
that it seems to mean very different
that it seems to mean very different things to different people some people
things to different people some people
things to different people some people when they hear the term Big Data they
when they hear the term Big Data they
when they hear the term Big Data they think of money they think about how it
think of money they think about how it
think of money they think about how it can be used by companies to design
can be used by companies to design
can be used by companies to design products more efficiently and market
products more efficiently and market
products more efficiently and market them more effectively other people when
them more effectively other people when
them more effectively other people when they hear the term Big Data they think
they hear the term Big Data they think
they hear the term Big Data they think about saving the world they think about
about saving the world they think about
about saving the world they think about how governments and NGOs can use this
how governments and NGOs can use this
how governments and NGOs can use this new kind of information to help solve
new kind of information to help solve
new kind of information to help solve important social problems some other
important social problems some other
important social problems some other people though when they hear the term
people though when they hear the term
people though when they hear the term Big Data they think of surveillance they
Big Data they think of surveillance they
Big Data they think of surveillance they think about all the enormous amounts of
think about all the enormous amounts of
think about all the enormous amounts of potentially very sensitive information
potentially very sensitive information
potentially very sensitive information that’s being collected about all of us
that’s being collected about all of us
that’s being collected about all of us all the time but for me when I think
all the time but for me when I think
all the time but for me when I think about big data I think about something
about big data I think about something
about big data I think about something different I think about a urinal so this
different I think about a urinal so this
different I think about a urinal so this though as you all recognize is not just
though as you all recognize is not just
though as you all recognize is not just any urinal this is a very very special
any urinal this is a very very special
any urinal this is a very very special urinal this is fountain by Duchamp this
urinal this is fountain by Duchamp this
urinal this is fountain by Duchamp this is one of my favorite pieces of art and
is one of my favorite pieces of art and
is one of my favorite pieces of art and what I love about this is the way that
what I love about this is the way that
what I love about this is the way that Duchamp took something that was created
Duchamp took something that was created
Duchamp took something that was created for one purpose and then repurposed it
for one purpose and then repurposed it
for one purpose and then repurposed it for something else and it turns out that
for something else and it turns out that
for something else and it turns out that this idea of repurposing is central to
this idea of repurposing is central to
this idea of repurposing is central to the way that scientists learn from big
the way that scientists learn from big
the way that scientists learn from big data because most of the big data
data because most of the big data
data because most of the big data sources that exist were created by
sources that exist were created by
sources that exist were created by companies and governments for reasons
companies and governments for reasons
companies and governments for reasons other than doing research and so
other than doing research and so
other than doing research and so scientists have to take data that was
scientists have to take data that was
scientists have to take data that was created for one purpose and like Duchamp
created for one purpose and like Duchamp
created for one purpose and like Duchamp repurpose it for something else and this
repurpose it for something else and this
repurpose it for something else and this is actually really different than the
is actually really different than the
is actually really different than the way social scientists have gone about
way social scientists have gone about
way social scientists have gone about learning about the world in the past and
learning about the world in the past and
learning about the world in the past and if I had to pick one piece of art that
if I had to pick one piece of art that
if I had to pick one piece of art that could illustrate this sort of more
could illustrate this sort of more
could illustrate this sort of more traditional style of social science
traditional style of social science
traditional style of social science research it would be David so when
research it would be David so when
research it would be David so when Michelangelo wanted to create David he
Michelangelo wanted to create David he
Michelangelo wanted to create David he didn’t look around for a piece of stone
didn’t look around for a piece of stone
didn’t look around for a piece of stone that kind of sort of looked like David
that kind of sort of looked like David
that kind of sort of looked like David he spent three years working to create
he spent three years working to create
he spent three years working to create David and so this illustrates really the
David and so this illustrates really the
David and so this illustrates really the two styles of research that we see right
two styles of research that we see right
two styles of research that we see right now there is a ready-made style where
now there is a ready-made style where
now there is a ready-made style where researchers Duchamp like take data that
researchers Duchamp like take data that
researchers Duchamp like take data that was created let’s say by Twitter or
was created let’s say by Twitter or
was created let’s say by Twitter or Facebook and repurpose it for research
Facebook and repurpose it for research
Facebook and repurpose it for research and then there’s the other style where
and then there’s the other style where
and then there’s the other style where researchers create their own data using
researchers create their own data using
researchers create their own data using like surveys or experiments but
like surveys or experiments but
like surveys or experiments but increasingly what we’re going to see is
increasingly what we’re going to see is
increasingly what we’re going to see is that there are limits to both of these
that there are limits to both of these
that there are limits to both of these pure approaches the Duchamp’s will
pure approaches the Duchamp’s will
pure approaches the Duchamp’s will discover that there are things that you
discover that there are things that you
discover that there are things that you can’t learn just by repurposing data and
can’t learn just by repurposing data and
can’t learn just by repurposing data and the michelangelo’s will increasingly
the michelangelo’s will increasingly
the michelangelo’s will increasingly feel like they’re being inefficient and
feel like they’re being inefficient and
feel like they’re being inefficient and wasteful by not taking advantage of all
wasteful by not taking advantage of all
wasteful by not taking advantage of all the ready-made big data that exists so I
the ready-made big data that exists so I
the ready-made big data that exists so I think the future of social research is a
think the future of social research is a
think the future of social research is a hybrid of these two strategies it’s
hybrid of these two strategies it’s
hybrid of these two strategies it’s going to be part ready-made and part
going to be part ready-made and part
going to be part ready-made and part custom-made and that’s going to allow us
custom-made and that’s going to allow us
custom-made and that’s going to allow us to do things that we could never do
to do things that we could never do
to do things that we could never do before with either of these pure
before with either of these pure
before with either of these pure strategies so let me give you an example
strategies so let me give you an example
strategies so let me give you an example of a beautiful piece of research that
of a beautiful piece of research that
of a beautiful piece of research that illustrates this hybrid strategy this is
illustrates this hybrid strategy this is
illustrates this hybrid strategy this is a study that was conducted by Josh
a study that was conducted by Josh
a study that was conducted by Josh blumenstock and colleagues and it was
blumenstock and colleagues and it was
blumenstock and colleagues and it was published in science a few years ago
published in science a few years ago
published in science a few years ago before I tell you about what they did
before I tell you about what they did
before I tell you about what they did let me tell you a little bit about the
let me tell you a little bit about the
let me tell you a little bit about the problem they were trying to solve let’s
problem they were trying to solve let’s
problem they were trying to solve let’s say you want to eliminate poverty in the
say you want to eliminate poverty in the
say you want to eliminate poverty in the world so that’s a goal I think many of
world so that’s a goal I think many of
world so that’s a goal I think many of us would share but one of the many many
us would share but one of the many many
us would share but one of the many many many problems you would encounter on the
many problems you would encounter on the
many problems you would encounter on the way to that goal is that it’s actually
way to that goal is that it’s actually
way to that goal is that it’s actually hard to know where poverty is and how
hard to know where poverty is and how
hard to know where poverty is and how poverty is changing over time then this
poverty is changing over time then this
poverty is changing over time then this may be kind of a surprising thing for
may be kind of a surprising thing for
may be kind of a surprising thing for people in wealthy countries because we
people in wealthy countries because we
people in wealthy countries because we happen to live in countries where for
happen to live in countries where for
happen to live in countries where for example in the US the National
example in the US the National
example in the US the National Statistical system collects reliable and
Statistical system collects reliable and
Statistical system collects reliable and accurate information about our society
accurate information about our society
accurate information about our society but in many of the world’s poorest
but in many of the world’s poorest
but in many of the world’s poorest countries where this information is most
countries where this information is most
countries where this information is most needed it’s not collected and so this is
needed it’s not collected and so this is
needed it’s not collected and so this is a challenge for policymakers because
a challenge for policymakers because
a challenge for policymakers because they’re essentially trying to operate in
they’re essentially trying to operate in
they’re essentially trying to operate in the dark they they can’t know where the
the dark they they can’t know where the
the dark they they can’t know where the problem is which makes it hard to target
problem is which makes it hard to target
problem is which makes it hard to target the resources effectively and it’s hard
the resources effectively and it’s hard
the resources effectively and it’s hard to see how well they’re doing because
to see how well they’re doing because
to see how well they’re doing because they can’t measure the outcomes and so
they can’t measure the outcomes and so
they can’t measure the outcomes and so many many people have worked on the
many many people have worked on the
many many people have worked on the problem of improving measurement of
problem of improving measurement of
problem of improving measurement of poverty but Josh blumenstock and
poverty but Josh blumenstock and
poverty but Josh blumenstock and colleagues took a different approach so
colleagues took a different approach so
colleagues took a different approach so what they did is they started with the
what they did is they started with the
what they did is they started with the call records from the largest mobile
call records from the largest mobile
call records from the largest mobile phone provider in Rwanda so this was 1.5
phone provider in Rwanda so this was 1.5
phone provider in Rwanda so this was 1.5 million customers and they have very
million customers and they have very
million customers and they have very detailed records about the patterns of
detailed records about the patterns of
detailed records about the patterns of calling and texting for all of these
calling and texting for all of these
calling and texting for all of these people but these ready-made call records
people but these ready-made call records
people but these ready-made call records do not have the thing that the research
do not have the thing that the research
do not have the thing that the research actually cared about which is a measure
actually cared about which is a measure
actually cared about which is a measure of poverty and so what they did is they
of poverty and so what they did is they
of poverty and so what they did is they took us
took us
took us sample a random sample from these set of
sample a random sample from these set of
sample a random sample from these set of call records they phoned the people up
call records they phoned the people up
call records they phoned the people up and they gave them a traditional social
and they gave them a traditional social
and they gave them a traditional social science survey to measure their poverty
science survey to measure their poverty
science survey to measure their poverty so they have this custom-made data set
so they have this custom-made data set
so they have this custom-made data set and they have this ready-made data set
and they have this ready-made data set
and they have this ready-made data set and then what I love about this study is
and then what I love about this study is
and then what I love about this study is the way that they link them up together
the way that they link them up together
the way that they link them up together so let me walk you through that process
so let me walk you through that process
so let me walk you through that process now so the first thing they did is a
now so the first thing they did is a
now so the first thing they did is a process called feature engineering where
process called feature engineering where
process called feature engineering where they created an enormous matrix where
they created an enormous matrix where
they created an enormous matrix where there’s one row for each person and one
there’s one row for each person and one
there’s one row for each person and one column for each feature and these
column for each feature and these
column for each feature and these features could be things like number of
features could be things like number of
features could be things like number of outgoing calls number of incoming calls
outgoing calls number of incoming calls
outgoing calls number of incoming calls but they could also be much more subtle
but they could also be much more subtle
but they could also be much more subtle so for example of the people that you’ve
so for example of the people that you’ve
so for example of the people that you’ve called what’s the probability if they
called what’s the probability if they
called what’s the probability if they called each other so they build many
called each other so they build many
called each other so they build many many complicated features like this and
many complicated features like this and
many complicated features like this and then they train a machine learning model
then they train a machine learning model
then they train a machine learning model to use those features to predict what
to use those features to predict what
to use those features to predict what people said on the survey so for example
people said on the survey so for example
people said on the survey so for example this machine learning model might
this machine learning model might
this machine learning model might discover that people who make more
discover that people who make more
discover that people who make more international calls tend to be more
international calls tend to be more
international calls tend to be more wealthy as an example of the kind of
wealthy as an example of the kind of
wealthy as an example of the kind of thing it might find it actually finds
thing it might find it actually finds
thing it might find it actually finds many more subtle patterns so once you
many more subtle patterns so once you
many more subtle patterns so once you have this machine learning model then
have this machine learning model then
have this machine learning model then you can impute the survey responses for
you can impute the survey responses for
you can impute the survey responses for all 1.49 nine million other people so
all 1.49 nine million other people so
all 1.49 nine million other people so impute is just a fancy social science
impute is just a fancy social science
impute is just a fancy social science term for guessing and so in essence what
term for guessing and so in essence what
term for guessing and so in essence what they’ve done here is by doing a survey
they’ve done here is by doing a survey
they’ve done here is by doing a survey of a thousand people they’ve been able
of a thousand people they’ve been able
of a thousand people they’ve been able to approximate what would happen if you
to approximate what would happen if you
to approximate what would happen if you did a survey of 1.5 million people then
did a survey of 1.5 million people then
did a survey of 1.5 million people then what they did is they estimated where
what they did is they estimated where
what they did is they estimated where everyone lived even though the call
everyone lived even though the call
everyone lived even though the call records did not have people’s names and
records did not have people’s names and
records did not have people’s names and addresses it turns out that what towers
addresses it turns out that what towers
addresses it turns out that what towers you make phone calls from at night is a
you make phone calls from at night is a
you make phone calls from at night is a pretty good approximation of where you
pretty good approximation of where you
pretty good approximation of where you live so here you see how they were able
live so here you see how they were able
live so here you see how they were able to connect up this ready-made call data
to connect up this ready-made call data
to connect up this ready-made call data with this custom-made survey data and
with this custom-made survey data and
with this custom-made survey data and now they’re able to produce high
now they’re able to produce high
now they’re able to produce high resolution Geographic estimates of
resolution Geographic estimates of
resolution Geographic estimates of poverty in Rwanda and here’s what those
poverty in Rwanda and here’s what those
poverty in Rwanda and here’s what those look like so this is a partial map of
look like so this is a partial map of
look like so this is a partial map of Rwanda I want to call attention to the
Rwanda I want to call attention to the
Rwanda I want to call attention to the square in the lower right corner so the
square in the lower right corner so the
square in the lower right corner so the scale of that is one kilometer by one
scale of that is one kilometer by one
scale of that is one kilometer by one kilometer and so you can see that they
kilometer and so you can see that they
kilometer and so you can see that they are able to make estimates for very very
are able to make estimates for very very
are able to make estimates for very very small geographic areas estimates that
small geographic areas estimates that
small geographic areas estimates that would be very helpful to policymakers so
would be very helpful to policymakers so
would be very helpful to policymakers so now you may be
now you may be
now you may be like are these estimates even accurate
like are these estimates even accurate
like are these estimates even accurate that’s the important question and the
that’s the important question and the
that’s the important question and the unfortunate truth is we don’t know
unfortunate truth is we don’t know
unfortunate truth is we don’t know because no one has ever made estimates
because no one has ever made estimates
because no one has ever made estimates this precise before however what you
this precise before however what you
this precise before however what you would really like to do is you would
would really like to do is you would
would really like to do is you would like to be able to compare these
like to be able to compare these
like to be able to compare these measurements somehow to some kind of
measurements somehow to some kind of
measurements somehow to some kind of gold standard measure and it turns out
gold standard measure and it turns out
gold standard measure and it turns out that a few years before their study the
that a few years before their study the
that a few years before their study the Rwandan government conducted a
Rwandan government conducted a
Rwandan government conducted a demographic and health survey which is a
demographic and health survey which is a
demographic and health survey which is a traditional expensive high-quality
traditional expensive high-quality
traditional expensive high-quality Social Survey that can reasonably be
Social Survey that can reasonably be
Social Survey that can reasonably be considered a gold standard so the
considered a gold standard so the
considered a gold standard so the demographic and health survey was not
demographic and health survey was not
demographic and health survey was not able to produce estimates at this small
able to produce estimates at this small
able to produce estimates at this small of a geographic level but it was able to
of a geographic level but it was able to
of a geographic level but it was able to produce estimates for the Rwanda sturdy
produce estimates for the Rwanda sturdy
produce estimates for the Rwanda sturdy regions and so what blumenstock and
regions and so what blumenstock and
regions and so what blumenstock and colleagues did is they aggregated their
colleagues did is they aggregated their
colleagues did is they aggregated their estimates up to these 30 regions and
estimates up to these 30 regions and
estimates up to these 30 regions and they produced a map like this so on this
they produced a map like this so on this
they produced a map like this so on this map darker colors indicate higher levels
map darker colors indicate higher levels
map darker colors indicate higher levels of poverty and so these are the
of poverty and so these are the
of poverty and so these are the estimates that come from this
estimates that come from this
estimates that come from this combination of the ready-made and
combination of the ready-made and
combination of the ready-made and custom-made data and now you may be
custom-made data and now you may be
custom-made data and now you may be wondering how well does this match the
wondering how well does this match the
wondering how well does this match the gold standard measure and so here’s the
gold standard measure and so here’s the
gold standard measure and so here’s the other estimate from the demographic and
other estimate from the demographic and
other estimate from the demographic and I’ll survive and so you can see in the
I’ll survive and so you can see in the
I’ll survive and so you can see in the paper they actually have a lot of very
paper they actually have a lot of very
paper they actually have a lot of very detailed technical comparisons which I’m
detailed technical comparisons which I’m
detailed technical comparisons which I’m just going to summarize the saying they
just going to summarize the saying they
just going to summarize the saying they match really well so for example you can
match really well so for example you can
match really well so for example you can see in the center of the countries both
see in the center of the countries both
see in the center of the countries both approaches detect the region of high
approaches detect the region of high
approaches detect the region of high poverty and then directly beneath that
poverty and then directly beneath that
poverty and then directly beneath that they have a region of higher levels of
they have a region of higher levels of
they have a region of higher levels of wealth so this is really exciting that
wealth so this is really exciting that
wealth so this is really exciting that this new approach to measurement can
this new approach to measurement can
this new approach to measurement can reproduce this much more traditional
reproduce this much more traditional
reproduce this much more traditional well-established measurement technique
well-established measurement technique
well-established measurement technique but then you might be saying well why
but then you might be saying well why
but then you might be saying well why should we care like if we already have
should we care like if we already have
should we care like if we already have this gold standard technique why should
this gold standard technique why should
this gold standard technique why should we be interested in this new technique
we be interested in this new technique
we be interested in this new technique at all and the answer is that this new
at all and the answer is that this new
at all and the answer is that this new technique is 10 times faster and 50
technique is 10 times faster and 50
technique is 10 times faster and 50 times cheaper so 50 times cheaper that
times cheaper so 50 times cheaper that
times cheaper so 50 times cheaper that is not like 10% cheaper 20% cheaper
is not like 10% cheaper 20% cheaper
is not like 10% cheaper 20% cheaper that’s like 50 times cheaper and so that
that’s like 50 times cheaper and so that
that’s like 50 times cheaper and so that makes a huge huge difference so that
makes a huge huge difference so that
makes a huge huge difference so that enables all kinds of new things so let
enables all kinds of new things so let
enables all kinds of new things so let me just give you one concrete example so
me just give you one concrete example so
me just give you one concrete example so instead of doing the demographic and
instead of doing the demographic and
instead of doing the demographic and health survey every five years as
health survey every five years as
health survey every five years as happens now you could do it every month
happens now you could do it every month
happens now you could do it every month so in wealthy countries we don’t have to
so in wealthy countries we don’t have to
so in wealthy countries we don’t have to wait every five years to get measures of
wait every five years to get measures of
wait every five years to get measures of our society we track things
our society we track things
our society we track things like the unemployment rate every month
like the unemployment rate every month
like the unemployment rate every month and why should it be that people in poor
and why should it be that people in poor
and why should it be that people in poor countries don’t have the same level of
countries don’t have the same level of
countries don’t have the same level of measurement as well so that’s just one
measurement as well so that’s just one
measurement as well so that’s just one example of the kind of thing that we can
example of the kind of thing that we can
example of the kind of thing that we can do when the cost drops by a factor of 50
do when the cost drops by a factor of 50
do when the cost drops by a factor of 50 and scientists will think of lots of
and scientists will think of lots of
and scientists will think of lots of other creative things that can be done
other creative things that can be done
other creative things that can be done as well now the scientist in me also
as well now the scientist in me also
as well now the scientist in me also wants to clarify that this is not a
wants to clarify that this is not a
wants to clarify that this is not a hundred percent of a fair apples to
hundred percent of a fair apples to
hundred percent of a fair apples to apples comparison because there are
apples comparison because there are
apples comparison because there are about 70 years of theoretical results
about 70 years of theoretical results
about 70 years of theoretical results and empirical know-how underlying the
and empirical know-how underlying the
and empirical know-how underlying the custom-made demographic and health
custom-made demographic and health
custom-made demographic and health survey we know a lot about that kind of
survey we know a lot about that kind of
survey we know a lot about that kind of technique and there’s much less of a
technique and there’s much less of a
technique and there’s much less of a theoretical foundation for this new
theoretical foundation for this new
theoretical foundation for this new approach that combines ready-mades and
approach that combines ready-mades and
approach that combines ready-mades and custom aids however learning from big
custom aids however learning from big
custom aids however learning from big data sources and figuring out how we can
data sources and figuring out how we can
data sources and figuring out how we can do it reliably is an incredibly active
do it reliably is an incredibly active
do it reliably is an incredibly active area of research right now and I’m
area of research right now and I’m
area of research right now and I’m really confident that we’re gonna make a
really confident that we’re gonna make a
really confident that we’re gonna make a lot of progress in the coming years so
lot of progress in the coming years so
lot of progress in the coming years so stepping back for a second from this
stepping back for a second from this
stepping back for a second from this study I hope that you can see now some
study I hope that you can see now some
study I hope that you can see now some of the points that we talked about at
of the points that we talked about at
of the points that we talked about at the beginning so this study is an
the beginning so this study is an
the beginning so this study is an example of how companies are potentially
example of how companies are potentially
example of how companies are potentially sitting on lots of valuable information
sitting on lots of valuable information
sitting on lots of valuable information that they are not currently extracting
that they are not currently extracting
that they are not currently extracting from their big data it’s also an example
from their big data it’s also an example
from their big data it’s also an example of how when analyzed appropriately big
of how when analyzed appropriately big
of how when analyzed appropriately big data sources can help address important
data sources can help address important
data sources can help address important social problems and it’s a reminder that
social problems and it’s a reminder that
social problems and it’s a reminder that enormous amounts of data are being
enormous amounts of data are being
enormous amounts of data are being collected about us but what I love about
collected about us but what I love about
collected about us but what I love about this study well I think it’s so great is
this study well I think it’s so great is
this study well I think it’s so great is because it’s a beautiful example of this
because it’s a beautiful example of this
because it’s a beautiful example of this hybrid style that is the ready-made call
hybrid style that is the ready-made call
hybrid style that is the ready-made call data alone was not enough the
data alone was not enough the
data alone was not enough the custom-made survey data was not enough
custom-made survey data was not enough
custom-made survey data was not enough it was only through the combination of
it was only through the combination of
it was only through the combination of these two things that they were able to
these two things that they were able to
these two things that they were able to accomplish this goal and so I hope that
accomplish this goal and so I hope that
accomplish this goal and so I hope that the next time you hear the phrase big
the next time you hear the phrase big
the next time you hear the phrase big data this will be what you think of
data this will be what you think of
data this will be what you think of because this is a future of social
because this is a future of social
because this is a future of social research in the digital age thank you
”
Be First to Comment