Part IV
CS229 Lecture Notes
Andrew Ng
Generative Learning algorithms
So far, we’ve mainly been talking about learning algorithms that model p(y|x;θ), the conditional distribution of y given x. For instance, logistic regression modeled p(y|x; θ) as hθ(x) = g(θT x) where g is the sigmoid func- tion. In these notes, we’ll talk about a different type of learning algorithm.
Consider a classification problem in which we want to learn to distinguish between elephants (y = 1) and dogs (y = 0), based on some features of an animal. Given a training set, an algorithm like logistic regression or the perceptron algorithm (basically) tries to find a straight line—that is, a decision boundary—that separates the elephants and dogs. Then, to classify a new animal as either an elephant or a dog, it checks on which side of the decision boundary it falls, and makes its prediction accordingly.
Here’s a different approach. First, looking at elephants, we can build a model of what elephants look like. Then, looking at dogs, we can build a separate model of what dogs look like. Finally, to classify a new animal, we can match the new animal against the elephant model, and match it against the dog model, to see whether the new animal looks more like the elephants or more like the dogs we had seen in the training set.
Algorithms that try to learn p(y|x) directly (such as logistic regression), or algorithms that try to learn mappings directly from the space of inputs X to the labels {0, 1}, (such as the perceptron algorithm) are called discrim- inative learning algorithms. Here, we’ll talk about algorithms that instead try to model p(x|y) (and p(y)). These algorithms are called generative learning algorithms. For instance, if y indicates whether an example is a dog (0) or an elephant (1), then p(x|y = 0) models the distribution of dogs’ features, and p(x|y = 1) models the distribution of elephants’ features.
After modeling p(y) (called the class priors) and p(x|y), our algorithm 1
2 can then use Bayes rule to derive the posterior distribution on y given x:
p(y|x) = p(x|y)p(y). p(x)
Here, the denominator is given by p(x) = p(x|y = 1)p(y = 1) + p(x|y = 0)p(y = 0) (you should be able to verify that this is true from the standard properties of probabilities), and thus can also be expressed in terms of the quantities p(x|y) and p(y) that we’ve learned. Actually, if were calculating p(y|x) in order to make a prediction, then we don’t actually need to calculate the denominator, since
arg max p(y|x) = arg max p(x|y)p(y) y yp(x)
= arg max p(x|y)p(y). y
1 Gaussian discriminant analysis
The first generative learning algorithm that we’ll look at is Gaussian discrim- inant analysis (GDA). In this model, we’ll assume that p(x|y) is distributed according to a multivariate normal distribution. Let’s talk briefly about the properties of multivariate normal distributions before moving on to the GDA model itself.
1.1 The multivariate normal distribution
The multivariate normal distribution in d-dimensions, also called the multi- variate Gaussian distribution, is parameterized by a mean vector μ ∈ Rd and a covariance matrix Σ ∈ Rd×d, where Σ ≥ 0 is symmetric and positive semi-definite. Also written “N (μ, Σ)”, its density is given by:
p(x;μ,Σ) = 1 exp−1(x−μ)TΣ−1(x−μ). (2π)d/2|Σ|1/2 2
In the equation above, “|Σ|” denotes the determinant of the matrix Σ.
For a random variable X distributed N(μ,Σ), the mean is (unsurpris-
ingly) given by μ:
E[X] = xp(x;μ,Σ)dx = μ
x
The covariance of a vector-valued random variable Z is defined as Cov(Z) = E[(Z − E[Z])(Z − E[Z])T ]. This generalizes the notion of the variance of a
3
real-valued random variable. The covariance can also be defined as Cov(Z) = E[ZZT ] − (E[Z])(E[Z])T . (You should be able to prove to yourself that these two definitions are equivalent.) If X ∼ N (μ, Σ), then
Cov(X) = Σ.
Here are some examples of what the density of a Gaussian distribution
looks like:
0.25
0.2
0.15
0.1
0.05
3
2
3
12 01
−1 0
−2
−3 −3
−1 −2
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
33 22
33 1212
0101 −1 0 −1 0
−2
−3 −3
−1
−2
−1
The left-most figure shows a Gaussian with mean zero (that is, the 2×1 zero-vector) and covariance matrix Σ = I (the 2×2 identity matrix). A Gaus- sian with zero mean and identity covariance is also called the standard nor- mal distribution. The middle figure shows the density of a Gaussian with zero mean and Σ = 0.6I; and in the rightmost figure shows one with , Σ = 2I. We see that as Σ becomes larger, the Gaussian becomes more “spread-out,” and as it becomes smaller, the distribution becomes more “compressed.”
Let’s look at some more examples.
0.25 0.25 0.2 0.2 0.15 0.15 0.1 0.1 0.05 0.05
33 22
11
00 33
−1 2 −1 2 11
−2 0 −2 0
−2
−2
−3
−3
0.25 0.2 0.15 0.1 0.05
3
2
1
0
−1 2 1
−2 0 −1
−3
−2
−3
3
−3 −2 −3
−3
−2
The figures above show Gaussians with mean 0, and with covariance matrices respectively
Σ=1 0; Σ= 1 0.5; Σ= 1 0.8. 0 1 0.5 1 0.8 1
The leftmost figure shows the familiar standard normal distribution, and we see that as we increase the off-diagonal entry in Σ, the density becomes more
−1
−1
−3
4
“compressed” towards the 45◦ line (given by x1 = x2). We can see this more clearly when we look at the contours of the same three densities:
3
2
1
0
−1
−2
−3
−3 −2 −1 0 1 2 3
Here’s one last set of examples generated by varying Σ:
33
22
11
00
−1 −1
−2 −2
−3 −3
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3
The plots above used, respectively,
Σ=1 -0.5;Σ=1 -0.8;Σ=3 0.8. -0.5 1 -0.8 1 0.8 1
From the leftmost and middle figures, we see that by decreasing the off- diagonal elements of the covariance matrix, the density now becomes “com- pressed” again, but in the opposite direction. Lastly, as we vary the pa- rameters, more generally the contours will form ellipses (the rightmost figure showing an example).
As our last set of examples, fixing Σ = I, by varying μ, we can also move the mean of the density around.
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
33 22
33 1212
0101 −1 0 −1 0
3
2
1
0
−1
−2
−3
−3 −2 −1 0 1 2 3
3
2
1
0
−1
−2
−3
−3 −2 −1 0 1 2 3
3
2
1
0
−1
−2
−3
−3 −2 −1 0 1 2 3
0.25
0.2
0.15
0.1
0.05
3
2
3
12 01
−1 0
−2
−3 −3
−1 −2
−2
−3 −3
−1
−2
−1
−2
−2
−3
−3
The figures above were generated using Σ = I, and respectively μ=1; μ=-0.5; μ= -1 .
0 0 -1.5
1.2 The Gaussian Discriminant Analysis model
When we have a classification problem in which the input features x are continuous-valued random variables, we can then use the Gaussian Discrim- inant Analysis (GDA) model, which models p(x|y) using a multivariate nor- mal distribution. The model is:
y ∼ Bernoulli(φ) x|y=0 ∼ N(μ0,Σ) x|y=1 ∼ N(μ1,Σ)
Writing out the distributions, this is:
p(y) = p(x|y=0) =
p(x|y=1) =
Here, the parameters of our model are φ, Σ, μ0 and μ1. (Note that while there’re two different mean vectors μ0 and μ1, this model is usually applied using only one covariance matrix Σ.) The log-likelihood of the data is given by
φy(1 − φ)1−y 1
exp−1(x−μ0)TΣ−1(x−μ0) (2π)d/2|Σ|1/2 2
5
1
(2π)d/2|Σ|1/2 2
exp−1(x−μ1)TΣ−1(x−μ1)
l(φ, μ0, μ1, Σ) =
n
log p(x(i), y(i); φ, μ0, μ1, Σ)
i=1 n
= log p(x(i)|y(i); μ0, μ1, Σ)p(y(i); φ). i=1
6
By maximizing l with respect to the parameters, we find the maximum like- lihood estimate of the parameters (see problem set 1) to be:
1 n
φ=n 1{y(i)=1}
i=1
ni=1 1{y(i) = 0}x(i)
μ0 = ni=11{y(i)=0}
ni=1 1{y(i) = 1}x(i) μ =
1 ni=1 1{y(i) = 1} 1n
Σ = n (x(i) −μy(i))(x(i) −μy(i))T. i=1
Pictorially, what the algorithm is doing can be seen in as follows:
1
0
−1
−2
−3
−4
−5
−6
−7
−2 −1 0 1 2 3 4 5 6 7
Shown in the figure are the training set, as well as the contours of the two Gaussian distributions that have been fit to the data in each of the two classes. Note that the two Gaussians have contours that are the same shape and orientation, since they share a covariance matrix Σ, but they have different means μ0 and μ1. Also shown in the figure is the straight line giving the decision boundary at which p(y = 1|x) = 0.5. On one side of the boundary, we’ll predict y = 1 to be the most likely outcome, and on the other side, we’ll predict y = 0.
1.3 Discussion: GDA and logistic regression
The GDA model has an interesting relationship to logistic regression. If we view the quantity p(y = 1|x; φ, μ0, μ1, Σ) as a function of x, we’ll find that it can be expressed in the form
p(y = 1|x;φ,Σ,μ0,μ1) = 1 , 1 + exp(−θT x)
whereθissomeappropriatefunctionofφ,Σ,μ0,μ1.1 Thisisexactlytheform that logistic regression—a discriminative algorithm—used to model p(y = 1|x).
When would we prefer one model over another? GDA and logistic regres- sion will, in general, give different decision boundaries when trained on the same dataset. Which is better?
We just argued that if p(x|y) is multivariate gaussian (with shared Σ), then p(y|x) necessarily follows a logistic function. The converse, however, is not true; i.e., p(y|x) being a logistic function does not imply p(x|y) is multivariate gaussian. This shows that GDA makes stronger modeling as- sumptions about the data than does logistic regression. It turns out that when these modeling assumptions are correct, then GDA will find better fits to the data, and is a better model. Specifically, when p(x|y) is indeed gaus- sian (with shared Σ), then GDA is asymptotically efficient. Informally, this means that in the limit of very large training sets (large n), there is no algorithm that is strictly better than GDA (in terms of, say, how accurately they estimate p(y|x)). In particular, it can be shown that in this setting, GDA will be a better algorithm than logistic regression; and more generally, even for small training set sizes, we would generally expect GDA to better.
In contrast, by making significantly weaker assumptions, logistic regres- sion is also more robust and less sensitive to incorrect modeling assumptions. There are many different sets of assumptions that would lead to p(y|x) taking the form of a logistic function. For example, if x|y = 0 ∼ Poisson(λ0), and x|y = 1 ∼ Poisson(λ1), then p(y|x) will be logistic. Logistic regression will also work well on Poisson data like this. But if we were to use GDA on such data—and fit Gaussian distributions to such non-Gaussian data—then the results will be less predictable, and GDA may (or may not) do well.
To summarize: GDA makes stronger modeling assumptions, and is more data efficient (i.e., requires less training data to learn “well”) when the mod- eling assumptions are correct or at least approximately correct. Logistic
1This uses the convention of redefining the x(i)’s on the right-hand-side to be (d + 1)-
dimensional vectors by adding the extra coordinate x(i) = 1; see problem set 1. 0
7
8
regression makes weaker assumptions, and is significantly more robust to deviations from modeling assumptions. Specifically, when the data is in- deed non-Gaussian, then in the limit of large datasets, logistic regression will almost always do better than GDA. For this reason, in practice logistic re- gression is used more often than GDA. (Some related considerations about discriminative vs. generative models also apply for the Naive Bayes algo- rithm that we discuss next, but the Naive Bayes algorithm is still considered a very good, and is certainly also a very popular, classification algorithm.)
2 Naive Bayes
In GDA, the feature vectors x were continuous, real-valued vectors. Let’s now talk about a different learning algorithm in which the xj’s are discrete- valued.
For our motivating example, consider building an email spam filter using machine learning. Here, we wish to classify messages according to whether they are unsolicited commercial (spam) email, or non-spam email. After learning to do this, we can then have our mail reader automatically filter out the spam messages and perhaps place them in a separate mail folder. Classifying emails is one example of a broader set of problems called text classification.
Let’s say we have a training set (a set of emails labeled as spam or non- spam). We’ll begin our construction of our spam filter by specifying the features xj used to represent an email.
We will represent an email via a feature vector whose length is equal to the number of words in the dictionary. Specifically, if an email contains the j-th word of the dictionary, then we will set xj = 1; otherwise, we let xj = 0. For instance, the vector
10 aardvark 0 aardwolf
. . x=. .
1 buy ..
. .
0 zygmurgy
is used to represent an email that contains the words “a” and “buy,” but not
9
“aardvark,” “aardwolf” or “zygmurgy.”2 The set of words encoded into the feature vector is called the vocabulary, so the dimension of x is equal to the size of the vocabulary.
Having chosen our feature vector, we now want to build a generative model. So, we have to model p(x|y). But if we have, say, a vocabulary of 50000 words, then x ∈ {0, 1}50000 (x is a 50000-dimensional vector of 0’s and 1’s), and if we were to model x explicitly with a multinomial distribution over the 250000 possible outcomes, then we’d end up with a (250000 −1)-dimensional parameter vector. This is clearly too many parameters.
To model p(x|y), we will therefore make a very strong assumption. We will assume that the xi’s are conditionally independent given y. This assumption is called the Naive Bayes (NB) assumption, and the resulting algorithm is called the Naive Bayes classifier. For instance, if y = 1 means spam email; “buy” is word 2087 and “price” is word 39831; then we are assuming that if I tell you y = 1 (that a particular piece of email is spam), then knowledge of x2087 (knowledge of whether “buy” appears in the message) will have no effect on your beliefs about the value of x39831 (whether “price” appears). More formally, this can be written p(x2087|y) = p(x2087|y, x39831). (Note that this is not the same as saying that x2087 and x39831 are independent, which would have been written “p(x2087) = p(x2087|x39831)”; rather, we are only assuming that x2087 and x39831 are conditionally independent given y.)
We now have:
p(x1, . . . , x50000|y)
= p(x1|y)p(x2|y, x1)p(x3|y, x1, x2) · · · p(x50000|y, x1, . . . , x49999)
= p(x1|y)p(x2|y)p(x3|y) · · · p(x50000|y) d
= p(xj|y) j=1
The first equality simply follows from the usual properties of probabilities, and the second equality used the NB assumption. We note that even though
2Actually, rather than looking through an English dictionary for the list of all English words, in practice it is more common to look through our training set and encode in our feature vector only the words that occur at least once there. Apart from reducing the number of words modeled and hence reducing our computational and space requirements, this also has the advantage of allowing us to model/include as a feature many words that may appear in your email (such as “cs229”) but that you won’t find in a dictionary. Sometimes (as in the homework), we also exclude the very high frequency words (which will be words like “the,” “of,” “and”; these high frequency, “content free” words are called stop words) since they occur in so many documents and do little to indicate whether an email is spam or non-spam.
10
the Naive Bayes assumption is an extremely strong assumptions, the resulting algorithm works well on many problems.
Our model is parameterized by φj|y=1 = p(xj = 1|y = 1), φj|y=0 = p(xj = 1|y = 0), and φy = p(y = 1). As usual, given a training set {(x(i),y(i));i = 1, . . . , n}, we can write down the joint likelihood of the data:
n
L(φy, φj|y=0, φj|y=1) = p(x(i), y(i)).
i=1
Maximizing this with respect to φy,φj|y=0 and φj|y=1 gives the maximum
likelihood estimates:
φy =
ni=1 1{y(i) = 1} n
n 1{x(i) =1∧y(i) =1} i=1 j
φj|y=1
φj|y=0 = ni=1 1{y(i) = 0}
= ni=1 1{y(i) = 1}
n 1{x(i) =1∧y(i) =0}
i=1 j
In the equations above, the “∧” symbol means “and.” The parameters have a very natural interpretation. For instance, φj|y=1 is just the fraction of the spam (y = 1) emails in which word j does appear.
Having fit all these parameters, to make a prediction on a new example with features x, we then simply calculate
p(x|y = 1)p(y = 1) p(x)
p(y = 1|x) =
=d d ,
and pick whichever class has the higher posterior probability.
Lastly, we note that while we have developed the Naive Bayes algorithm mainly for the case of problems where the features xj are binary-valued, the generalizationtowherexj cantakevaluesin{1,2,…,kj}isstraightforward. Here, we would simply model p(xj |y) as multinomial rather than as Bernoulli. Indeed, even if some original input attribute (say, the living area of a house, as in our earlier example) were continuous valued, it is quite common to discretize it—that is, turn it into a small set of discrete values—and apply Naive Bayes. For instance, if we use some feature xj to represent living area,
we might discretize the continuous values as follows:
dj=1 p(xj |y = 1) p(y = 1)
j=1p(xj|y=1) p(y=1)+ j=1p(xj|y=0) p(y=0)
11
Living area (sq. feet) < 400 400-800 800-1200 1200-1600 >1600 xi 12345
Thus, for a house with living area 890 square feet, we would set the value of the corresponding feature xj to 3. We can then apply the Naive Bayes algorithm, and model p(xj|y) with a multinomial distribution, as described previously. When the original, continuous-valued attributes are not well- modeled by a multivariate normal distribution, discretizing the features and using Naive Bayes (instead of GDA) will often result in a better classifier.
2.1 Laplace smoothing
The Naive Bayes algorithm as we have described it will work fairly well for many problems, but there is a simple change that makes it work much better, especially for text classification. Let’s briefly discuss a problem with the algorithm in its current form, and then talk about how we can fix it.
Consider spam/email classification, and let’s suppose that, we are in the year of 20xx, after completing CS229 and having done excellent work on the project, you decide around May 20xx to submit work you did to the NeurIPS conference for publication.3 Because you end up discussing the conference in your emails, you also start getting messages with the word “neurips” in it. But this is your first NeurIPS paper, and until this time, you had not previously seen any emails containing the word “neurips”; in particular “neurips” did not ever appear in your training set of spam/non-spam emails. Assuming that “neurips” was the 35000th word in the dictionary, your Naive Bayes spam filter therefore had picked its maximum likelihood estimates of the parameters φ35000|y to be
φ35000|y=1 φ35000|y=0
n
i=1 n
i=1
1{x(i) =1∧y(i) =1} 35000
= =
ni=1 1{y(i) = 1}
1{x(i) =1∧y(i) =0}
= 0
= 0
35000
ni=1 1{y(i) = 0}
I.e., because it has never seen “neurips” before in either spam or non-spam training examples, it thinks the probability of seeing it in either type of email is zero. Hence, when trying to decide if one of these messages containing
3NeurIPS is one of the top machine learning conferences. The deadline for submitting a paper is typically in May-June.
12 “neurips” is spam, it calculates the class posterior probabilities, and obtains
dj=1 p(xj|y = 1)p(y = 1)
p(y = 1|x) = dj=1 p(xj|y = 1)p(y = 1) + dj=1 p(xj|y = 0)p(y = 0)
= 0.
This is because each of the terms “dj=1 p(xj|y)” includes a term p(x35000|y) = 0 that is multiplied into it. Hence, our algorithm obtains 0/0, and doesn’t know how to make a prediction.
Stating the problem more broadly, it is statistically a bad idea to esti- mate the probability of some event to be zero just because you haven’t seen it before in your finite training set. Take the problem of estimating the mean of a multinomial random variable z taking values in {1, . . . , k}. We can pa- rameterize our multinomial with φj = p(z = j). Given a set of n independent observations {z(1), . . . , z(n)}, the maximum likelihood estimates are given by
φ j = ni = 1 1 { z ( i ) = j } . n
As we saw previously, if we were to use these maximum likelihood estimates, then some of the φj’s might end up as zero, which was a problem. To avoid this, we can use Laplace smoothing, which replaces the above estimate with
Here, we’ve added 1 to the numerator, and k to the denominator. Note that kj=1 φj = 1 still holds (check this yourself!), which is a desirable property since the φj’s are estimates for probabilities that we know must sum to 1. Also, φj ̸= 0 for all values of j, solving our problem of probabilities being estimated as zero. Under certain (arguably quite strong) conditions, it can be shown that the Laplace smoothing actually gives the optimal estimator of the φj’s.
Returning to our Naive Bayes classifier, with Laplace smoothing, we therefore obtain the following estimates of the parameters:
1 + ni=1 1{z(i) = j} φj = k+n .
φj|y=1 = φj|y=0 =
1+n 1{x(i) =1∧y(i) =1} i=1 j
n (i)
2+ i=11{y =1}
1+ n 1{x(i) =1∧y(i) =0} i=1 j
2 + ni=1 1{y(i) = 0}
13
(In practice, it usually doesn’t matter much whether we apply Laplace smooth- ing to φy or not, since we will typically have a fair fraction each of spam and non-spam messages, so φy will be a reasonable estimate of p(y = 1) and will be quite far from 0 anyway.)
2.2 Event models for text classification
To close off our discussion of generative learning algorithms, let’s talk about one more model that is specifically for text classification. While Naive Bayes as we’ve presented it will work well for many classification problems, for text classification, there is a related model that does even better.
In the specific context of text classification, Naive Bayes as presented uses the what’s called the Bernoulli event model (or sometimes multi-variate Bernoulli event model). In this model, we assumed that the way an email is generated is that first it is randomly determined (according to the class priors p(y)) whether a spammer or non-spammer will send you your next message. Then, the person sending the email runs through the dictionary, deciding whether to include each word j in that email independently and according to the probabilities p(xj = 1|y) = φj|y. Thus, the probability of a message was given by p(y) dj=1 p(xj |y).
Here’s a different model, called the Multinomial event model. To describe this model, we will use a different notation and set of features for representing emails. We let xj denote the identity of the j-th word in the email. Thus, xj is now an integer taking values in {1, . . . , |V |}, where |V | is the size of our vocabulary (dictionary). An email of d words is now rep- resented by a vector (x1,x2,…,xd) of length d; note that d can vary for different documents. For instance, if an email starts with “A NeurIPS . . . ,” then x1 = 1 (“a” is the first word in the dictionary), and x2 = 35000 (if “neurips” is the 35000th word in the dictionary).
In the multinomial event model, we assume that the way an email is generated is via a random process in which spam/non-spam is first deter- mined (according to p(y)) as before. Then, the sender of the email writes the email by first generating x1 from some multinomial distribution over words (p(x1|y)). Next, the second word x2 is chosen independently of x1 but from the same multinomial distribution, and similarly for x3, x4, and so on, until all d words of the email have been generated. Thus, the overall probability of a message is given by p(y)dj=1 p(xj|y). Note that this formula looks like the one we had earlier for the probability of a message under the Bernoulli event model, but that the terms in the formula now mean very different things. In particular xj|y is now a multinomial, rather than a Bernoulli distribution.
14
The parameters for our new model are φy = p(y) as before, φk|y=1 = p(xj =k|y=1)(foranyj)andφk|y=0 =p(xj =k|y=0). Notethatwehave assumed that p(xj|y) is the same for all values of j (i.e., that the distribution according to which a word is generated does not depend on its position j within the email).
If we are given a training set {(x(i),y(i));i = 1,…,n} where x(i) = (x(i),x(i),…,x(i))(here,d isthenumberofwordsinthei-trainingexample),
12di i
the likelihood of the data is given by
n
p(x(i), y(i))
L(φy, φk|y=0, φk|y=1) =
= p(x(i)|y; φk|y=0, φk|y=1)
i=1
n di
p(y(i); φy). Maximizing this yields the maximum likelihood estimates of the parameters:
j i=1 j=1
n di 1{x(i) =k∧y(i) =1} i=1 j=1 j
ni=1 1{y(i) = 1} φy= n .
φk|y=1 = φk|y=0 =
ni=1 1{y(i) = 1}di
n di (i) (i)
i=1j=11{xj =k∧y =0} ni=1 1{y(i) = 0}di
If we were to apply Laplace smoothing (which is needed in practice for good performance) when estimating φk|y=0 and φk|y=1, we add 1 to the numerators and |V | to the denominators, and obtain:
1+n di 1{x(i) =k∧y(i) =1} i=1 j=1 j
φk|y=1 = φk|y=0 =
While not necessarily the very best classification algorithm, the Naive Bayes classifier often works surprisingly well. It is often also a very good “first thing to try,” given its simplicity and ease of implementation.
|V | + ni=1 1{y(i) = 1}di 1+n di 1{x(i) =k∧y(i) =0}
i=1 j=1 j
|V|+n 1{y(i) =0}d .
i=1 i