Learn how to model social and economic networks and their impact on human behavior. How do networks form, why do they exhibit certain patterns, and how does their structure impact diffusion, learning, and other behaviors? We will bring together models and techniques from economics, sociology, math, physics, statistics and computer science to answer these questions. The course begins with some empirical background on social and economic networks, and an overview of concepts used to describe and measure networks. Next, we will cover a set of models of how networks form, including random network models as well as strategic formation models, and some hybrids. We will then discuss a series of models of how networks impact behavior, including contagion, diffusion, learning, and peer influences. You can find a more detailed syllabus here: http://web.stanford.edu/~jacksonm/Networks-Online-Syllabus.pdf You can find a short introductory videao here: http://web.stanford.edu/~jacksonm/Intro_Networks.mp4
5.2: Bass Model
Loading...
來自 Stanford University 的課程
Social and Economic Networks: Models and Analysis
333 個評分
從本節課中
Diffusion on Networks
Empirical Background, The Bass Model, Random Network Models of Contagion, The SIS model, Fitting a Simulated Model to Data.
與講師見面
Matthew O. JacksonProfessor
Economics
Hi, folks. We're back and we're talking more about
diffusion and, in particular, were going to start looking at one of the best
known models in the literature on diffusion, which is known as the Bass
Model. And it's been used quite extensively
especially in marketing and trying to understand things.
So, we've gone through some questions and background.
And were going to start with the Bass Model.
And one thing about the Bass Model is networks are not going to make on
explicit appearance here, were just going to have social interaction in the
background. And after we've looked at the Bass Model,
then we'll enrich the analysis by bringing interaction structure in
explicitly, and try to understand how things diffuse over time.
Okay, so why is the Bass Model interesting?
It's a benchmark model. It gives us very stark and simple
structure which can begin to produce things like the S shapes that we saw
before. And there's going to be two actions,
behaviors or states, so either you're not infected or you are, or you're not
adopting the new product or you are. and what the Bass Model will keep track
of is over time, what's the fraction of individuals who have adopted say, taken
action one by time t, okay? And in this model, you move from state 0
to state 1, you don't move back. So, everybody starts at state 0, then
people start moving towards state 1. Some fraction of the people can move to
state 1 and that will just move over time, and we'll just keep track over time
of how many people have now moved to state 1.
So, they have seen the movie, or they've adopted corn, or they caught the flu
etc., okay? So, it's a one-way street and we keep
track of that over time. Okay, two key parameters in the Bass
Model. p is going to be a rate of spontaneous
innovation or adoption. So, that's going to be a rate that's
independent of what else is going on in the economy or the world or the society.
So, some people will just go and decide to see a movie regardless of what's going
on, or they will adopt corn, regardless of whether other people are doing it, if
they've heard about it, etc. And there's some other individuals who
will actually do it as a function of having heard about it, or somehow
imitated other individuals who were doing it.
So, there's these two parameters p and q. And the Bass Model basically boils down
then to a, a simple equation which keeps track of the differential of the fraction
of people who have adopted at time t as, as t varies, okay?
So, how is this changing with time? Well, there's two parts to it.
So, at any point in time, there the people who haven't yet adopted, so these
are the people who have not yet adopted yet moved to one, right, so they haven't
become a one yet so they're still zeros. So, 1 minus Ft are the people who are
still 0 and can possibly change. And then, the fraction of those people
who change, well, some of them change directly spontaneously, that's the p.
And the others change by imitating the existing population.
So, you also, you have a chance of, of just spontaneously deciding to become a
1, or you also have a chance of imitating somebody in the population and that's
proportional to how many in the population are already ones.
So, this q parameter says that you, you have some rate of imitation.
These are the ones existing. This is your imitation and so this is the
probability that you end up adopting 1 due to some imitation.
This is the chance that you do it spontaneously and these are the fraction
of people who've not yet adopted and who could make that change.
So, that gives us an expression for the differential overtime.
Very simple and intuitive model in terms of its basic building blocks.
Okay. So, when we look at this if you want to
solve this expression, right, and then you can start with an initial condition
of F of 0 equals 0. If you solve that, you get a simple
equation for what F of t looks like. And it depends on, on p and q, obviously.
and so higher p's and q's are going to lead to faster diffusion so more people
adopting by any particular time, lower p's and ques are going to lead to lower
adoption rates. And so this thing will be increasing in p
and increasing in q at, at any point in time the number of people would adopt by
that time. Okay, so you get a simple solution.
And then, let's look at the aspects of this and, and why the model has been so
well used and, and is well-known. Okay, so first thing it's, it's going to
end up giving an s shape, if q is bigger than p is going to tend to 1 in the limit
as t becomes large. and basically what's going on in this
model is initially, only, the only way that people can adopt is, is mainly
going to be through p. And then eventually, q is going to become
the important parameter. and things will slow down as your, a
fraction of people that have adopted eventually reaches 1.
So, if we go back and look at exactly what this equation looks like, let's try
and analyze this in a little more detail and try and understand why we get
interesting dynamics out of this. Okay, so first of all when Ft gets close
to 1, this thing is going to have to slow down.
And so, what happens is that the grade at which you're gaining new people adopting
as F of t gets close to 1, this thing gets close to 0, right?
And so, this thing has to get close to 0. So, this thing is going to tend to 0 as
Ft gets close to 1. So eventually, it has to slow down just
because the fraction of people who haven't adopted yet becomes small, so
even if a lot of them are adopting, there's just not many of them left and
that's what gives you the last part of the curve.
So that's very intuitive. And it's coming directly out of the, the
limitation that this thing has to converge at most to 1.
When you're initially at 0, then this part is going to go to 0, right?
So, there's no imitation going on because there's nobody to imitate.
And everything is just happening from the spontaneous adopters.
So, initially, this thing is going to, this thing is going to look like 1,
everybodycan still adopt, but all of it's going to happen through the p.
So, what starts out is you start out with a slope of p, right, so you start out
with some initial adoption rate from 0, you're going to start out at a slope of
p, and then eventually, the q's going to start kicking in.
So, as you get more and more people adopting here, then the acute can kick
in, and that can begin to, to give you the S shape.
So, the idea can be that the S shape can start going up, as q begins to kick in,
as more people start imitating, okay? But there's, there's a competing factor,
which is, as more people are adopting, this thing is also going down, right, so
you have fewer people left to adopt. So, whether or not you get this S shape
is going to be a race between the increase in q and the decrease in the
population that's left. And eventually, we know it has to, to, to
asymptote and, and be concave. And so, the question is whether it's
going to be initially convex. And so, when we can, we can analyze that
by looking at this process close to 0. We know that the initial slope is, is p,
so let's look at, at what happens at some small epsilon.
So, we've, we've just started moving out. And we'll see whether we're going to
start accelerating or not. Is it accelerating or is it going to be
already decreasing in speed? So, what does dF dt look like at, at some
small epsilon? It's going to look like p plus q epsilon
times 1 minus epsilon. So, if you just plug in a small epsilon
for this, you get this. To a need to get the initial convexity,
you need this thing to be bigger than p, right?
We have to be accelerating, the, the slope started out at p, now we have to be
getting a slightly larger slope. So initially, to get that S shape, you're
going to have to have this be bigger than p, and what does that tell us?
that tells us basically that q is going to have to be bigger than p.
So, for very small epsilon, the only way that you can have this thing be bigger
than p is to have q be bigger than p. So, if q is bigger than p then you'll get
it effectively the q, q epsilon is going to be bigger than the p epsilon.
and so we end up with a situation where you get the convex initial condition.
So, q bigger than p gives us the initial growth where we get a convexity at the
beginning, okay? So, we get this S shape here if p is
bigger than q, we get initially the, the slope is p, then it starts accelerating.
And eventually, when F of t gets very large, it's going to have to slow down
because now there's just not many of the adopters left, and then things slow down
and we eventually get the asymptote towards to 1, okay?
So, that's the Bass Model. very compact, simple, easy model.
the the, this model has been used quite extensively.
Why it has been used so extensively? it's, it's very compact, so if you can
begin, you know, suppose that we've, we're just here at some time period,
that's already enough to estimate what p was and to begin to see, to estimate what
q is. So, as this process takes off, you don't
need much data to begin to analyze and form estimates of p and q.
Once you've got the estimates of p and q, you can get estimates of what the rest of
the process is going to look like. So, this model has been used extensively
in forecasting by trying to estimate from initial take up.
So, if you look, say, at the box office of a new movie, how many people go see
the movie in the first week? How many people go see the movie in the
second week? Based on that first week and second week,
you've already got a piece of two pieces of this curve.
You can begin to project how long is it going to take for this thing to keep you
know, to slow down and, and where will it eventually reach in terms of its
asymptote. So, the Bass Model has been enriched a
lot by adding in extra moving parts. Maybe some people wouldn't see the movie,
so here this assumes it goes up to 100%, maybe some people would never see a
certain movie. you can enrich it by, you know, adding in
different kinds of heterogeneity and so forth.
So, there's different ways to enrich this model.
And people use richer versions of the Bass Model for forecasting.
But this, this basic simple part gives us that S shape in a very simple way.
And we realize that, you know, it, it's this combination of social imitation,
which grows with time, which gives us that convexity, which is important here.
Okay, beyond the, the Bass Structure what we're going to look at next is component
structures and start bringing the network in.
So here, you know, more generally, the reach of diffusion is going to be bounded
by some component structure. So, if, if there's certain groups of
people that don't interact with others, then you know, things aren't going to
move across one group. it could be that some players or, or
nodes are immune. So, if we're talking about the flu,
people could be, have a vaccine against it and maybe not catch it.
Or it could be that, you know, there's certain individuals who just wouldn't go
see a certain kind of movies, that will never see a horror movie and, and so you
could think of them as being immune to certain kinds of diffusion.
it might be that some relationships between individuals that are transmitting
with higher probability than others. So, if we start looking at the network
structure, we're going to see that, that there could be some probabilistic
transmission. so the answers of questions of when we
get diffusion and its extent aren't answered so easily by the simple Bass
Model. And we'll have to enrich these kinds of
diffusion models in order to answer those, and so that's what we'll take a
look at next.
