6. World’s 1st
peer-to-peer
lending platform
in 2004
£2.5 billion
lent to date,
and our growth is
accelerating
246,000
people have taken
a Zopa loan
59,000
actively invest
through Zopa
7.
8.
9. What is PyMC3?
• Probabilistic Programming in Python
• At release stage – so ready for production
• Theano based
• Powerful sampling algorithms
• Powerful model syntax
• Recent improvements include Gaussian Processes and enhanced Variational
Inference
11. Who uses it?
• Used widely in academia and industry
• https://github.com/pymc-devs/pymc3/wiki/PyMC3-Testimonials
• https://scholar.google.de/scholar?hl=en&as_sdt=0,5&sciodt=0,5&cites=6936955228135731
011&scipsc=&authuser=1&q=&scisbd=1
12. What is a Bayesian approach?
• The Bayesian world-view interprets probability as a measure of believability in an event, that
is, how confident are we that an event will occur.
• The Frequentist approach/ view is – considers that probability is the long-run frequency of
events.
• This doesn’t make much sense for say Presidential elections!
• Bayesians interpret a probability as a measure of beliefs. This allows us all to have different
priors.
13. Are Frequentist methods wrong?
• NO
• Least squares regression, LASSO regression and expectation-maximization are all powerful
and fast in many areas.
• Bayesian methods complement these techniques by solving problems that these approaches
can’t
• Or by illuminating the underlying system with more flexible modelling.
14. Example – Let’s look at text message data
This data comes from Cameron Davidson-Pilon from his own text message
history. He wrote the examples and the book this talk is based on. It’s cited at the
end.
15. Example – Inferring text message data
- A Poisson random variable is a very appropriate model for this type of count
data.
- The math will be something like C_{i} ∼ Poisson(λ)
- We don’t know what the lambda is – what is it?
16. Example – Inferring text message data (continued)
- It looks like the rate is higher later in the observation period.
- We’ll represent this with a ‘switchpoint’ – it’s a bit like how we write a delta
function (we use a day which we call tau)
- λ={λ1 if t<τ
- {λ2 if t≥τ
17. Priors – Or beliefs
- We call alpha a hyper-parameter or parent variable. In literal terms, it is a
parameter that influences other parameters.
- Alternatively, we could have two priors – one for each λi -- EXERCISE
We are interested in inferring the unknown λs. To use Bayesian inference, we need
to assign prior probabilities to the different possible values of λ What would be
good prior probability distributions for λ1 and λ2?
Recall that λ can be any positive number. As we saw earlier, the exponential
distribution provides a continuous density function for positive numbers, so it
might be a good choice for modelling λi But recall that the exponential distribution
takes a parameter of its own, so we'll need to include that parameter in our
model. Let's call that parameter α.
λ1∼Exp(α)
λ2∼Exp(α)
18. Priors - Continued
- We don’t care what our prior distribution (or integral) for the unknown variables
looks like.
- It’s probably intractable – so needs a method to solve.
- And we care about the posterior distribution
- What about τ?
- Due to the noisiness of the data, it’s difficult to pick out a priori where τ
might have occurred. We’ll pick a uniform prior belief to every possible
day. This is equivalent to saying
τ ∼ DiscreteUniform(1,70)
- This implies that the P(τ=k) = 1/70
19. The philosophy of Probabilistic Programming:
Our first hammer PyMC3
Another way of thinking about this: unlike a traditional program, which only
runs in the forward directions, a probabilistic program is run in both the
forward and backward direction. It runs forward to compute the
consequences of the assumptions it contains about the world (i.e., the model
space it represents), but it also runs backward from the data to constrain the
possible explanations. In practice, many probabilistic programming systems
will cleverly interleave these forward and backward operations to efficiently
home in on the best explanations. -- Beau Cronin – sold a Probabilistic
Programming focused company to Salesforce
20. Let’s specify our variables
import pymc3 as pm import theano.tensor as tt
with pm.Model() as model:
alpha = 1.0/count_data.mean() # Recall count_data is the
# variable that holds our txt counts
lambda_1 = pm.Exponential("lambda_1", alpha)
lambda_2 = pm.Exponential("lambda_2", alpha)
tau = pm.DiscreteUniform("tau", lower = 0, upper=n_count_data - 1)
21. Let’s create the switchpoint and add in
observations
with model:
idx = np.arange(n_count_data) # Index
lambda_ = pm.math.switch(tau >= idx, lambda_1, lambda_2)
with model:
observation = pm.Poisson("obs", lambda_, observed=count_data)
All our variables so far are random variables. We aren’t fixing any variables yet.
The variable observation combines our data (count_data), with our proposed data-
generation schema, given by the variable lambda_, through the observed keyword.
22. Let’s learn something
with model:
trace = pm.sample(10000, tune=5000)
We can think of the above code as a learning step. The machinery we
use is called Markov Chain Monte Carlo (MCMC), which is a whole
workshop. Let’s consider it a magic trick that helps us solve these
complicated formula.
This technique returns thousands of random variables from the
posterior distributions of λ1,λ2 and τ.
We can plot a histogram of the random variables to see what the
posterior distributions look like.
On next slide, we collect the samples (called traces in the MCMC
literature) into histograms.
23. The trace code
lambda_1_samples = trace['lambda_1’]
lambda_2_samples = trace['lambda_2’]
tau_samples = trace['tau']
We’ll leave out the plotting code – you can check in
the notebooks.
25. Interpretation
• The Bayesian methodology returns a distribution.
• We now have distributions to describe the unknown λ1,λ2 and τ
• We can see that the plausible values for the parameters are: λ1 is around 18, and λ2 is
around 23. The posterior distributions of the two lambdas are clearly distinct, indicating
that it is indeed likely that there was a change in the user’s text-message behaviour.
• Our analysis also returned a distribution τ. It’s posterior distribution is discrete. We can see
that near day 45, there was a 50% chance that the user’s behaviour changed. This confirms
that a change occurred because had no changed occurred tau would be more spread out.
We see that only a few days make any sense as potential transition points.
26. Why would I want to sample from the posterior?
• Entire books are devoted to explaining why.
• We’ll muse the posterior samples to answer the following question: what is expected
number of texts at day t, 0≤t≤70? Recall that the expected value of a Poisson variable is
equal to it’s parameter λ. Therefore the question is equivalent to what is the expected value
of λ at time t?
• In our code, let i index samples from the posterior distributions. Given a day t, we average
over all possible λi for the day t, using λi = λ1,i if t < τi (that is, if the behaviour change has
not yet occurred), else we use λi = λ2,i
27. Analysis results
- Our analysis strongly supports believing the users’ behaviour did change.
Otherwise the two lambdas would be closer in value.
- The change was sudden rather than gradual – we see this from tau’s strongly
peaked posterior distribution.
- It turns out the 45th day was Christmas and the book author was moving cities.
28. We introduced Bayesian Methods
Bayesian methods are about beliefs
PyMC3 allows building generative models
We get uncertainty estimates for free
We can add domain knowledge in priors