Description
Principal Component Analysis
PCA is a very popular feature transformation technique that has nice theoretical properties and is still
surprisingly simple to implement. The central method is based on the Singular Value Decomposition, which
states that any matrix can be re-written as the product of three different matrices with useful properties:
X = USV >
1https://scikit-learn.org/
Figure 1: Expected output for PCA and k-Means on the 2D test dataset. The Normalized Mutual Information score should be close to 0.95.
For our purposes, the V matrix is the most interesting, because it contains as its columns the eigenvectors of
the covariance matrix of X, which as we showed in class, are exactly the vectors that point in the direction
of “highest variance.”
To actually solve for this decomposition, we can use the Numpy library, specifically
the function numpy.linalg.svd. Note 1: read the documentation for this function carefully, so that you
know what you need to pass in and what is being returned. Use this function to finish the implementation of
PCA.find components, which should simply store the return values in the appropriate data members.
Note
2: don’t forget to center the data (subtract out the mean) before passing it to the svd function, otherwise
your results may be different.
With the V matrix in hand, computing the actual feature transform is very simple. If X is the centered
data, and VL is the matrix consisting of the first L columns of V , we can compute the transformed data X0
as follows:
X0 = XVL
Use this equation to implement the PCA.transform method, making sure to center the data, and only use
the first L columns2 of V . For the final method you need to implement, we can leverage one of the special
properties of V , namely that V
> = V
−1
.
This means we can write the inverse transform as:
X00 = X0V
>
L
Implement this in PCA.inv transform to complete the class, making sure to “undo” the centering you did
in the original transform (add the mean back in)3
.
With all three of these methods finished, your PCA
implementation should be complete, and you can test out the plotting methods provided for you to visualize
any high dimensional dataset by projecting it down to the first two or three principal components.
You can
also see how much “information” is lost in these projections by looking at the reconstruction error, which is
the difference between the original data, and the data passed through a low-dimensional transform and then
re-projected back to the original space. This is implemented for you in PCA.reconstruction error.
k-Means clustering
k-Means clustering is a simple iterative method for finding clusters that alternates between assigning clusters
to centers, and recomputing center locations given assigned points. Formally, we can describe the process as
follows:
2This is easy to do using Numpy indexing. For a useful tutorial, see https://docs.scipy.org/doc/numpy/reference/arrays.
indexing.html
3Note that unless VL = V (we use all of the columns of V), the reconstructed X00 won’t necessarily exactly equal X. This
difference is what PCA.reconstruction error measures.
Data: number of clusters K, data {x
(i)}
N
i=1
Result: cluster centers, {c
(k)}
K
k=1
Initialize each c
(k)
to a randomly chosen x
(i)
while not done do
Assign each x
(i)
to the closest c
(k)
Set each c
(k)
to be the average of all the x
(i) assigned to it
If cluster assignments did not change, done
end
Implement this algorithm in KMeans.cluster. Note that detecting whether cluster assignments have
changed can be tricky if there are ties in the distances, so you may want to make use of the helper function
consistent ordering, or you may want to handle this in your own way.
Once you’ve implemented this method, the KMeans class should be complete. Typically, the cluster
centers are what we’re after in k-Means clustering, but these can be used in a variety of different ways. If
the purpose is feature reduction/transformation, we can use distance to the discovered clusters as a new
set of features.
This is how the provided 2D and 3D plotting functions work: they first find two/three
clusters, then transform the data into this new space, which can then be used as a visualization of the high
dimensional data.
Alternatively, if visualization isn’t the goal we can instead label each point with the closest cluster center,
and then use the cluster centers themselves as a kind of “summary” of the data. It’s difficult to come up
with good measures of clustering performance when all we have is the data itself, but thankfully for each of
the provided datasets we also have an actual class label (even though we ignored it when clustering).
One
way to measure cluster performance when we have access to this class label data is by measuring the Mutual
Information between the cluster labels and the class labels. If the clusters do a good job of summarizing the
data, then they should have a strong statistical relationship with the class labels.
The standard statistical
formula for measuring this kind of relationship is known as Mutual Information, and it’s defined to be
MI(U, V ) = X
i
X
j
p(Ui
, Vj ) · log p(Ui
, Vj )
p(Ui) · p(Vj )
where p(Ui
, Vj ) is the joint probability of a datapoint having both cluster label i and class label j, while
p(Ui) and p(Vj ) are the respective marginal probabilities for cluster label i and class label j.
The max
and min values of Mutual Information are dependent on the number of class and cluster labelings, so this
equation is often normalized to be a value between 0 and 1, with values near 0 having a weak relationship, and values near 1 having a strong relationship. This normalized version is implemented for you in
KMeans.normalized mutual information.
Try this out on the testing dataset with two clusters and you
should see a Normalized Mutual Information very close to 1, indicating that the cluster labels agree with
the class labels.
Questions
1. PCA
(a) Run PCA on the three datasets, iris, HTRU2, and digits. For each dataset, visualize the 2D
and 3D projections of the data (6 graphs). Make sure each point is colored according to its class
label (see the “with labels” example in the template main function).
(b) If we were to use PCA to transform the original features into those visualized in the previous
question, do you think it would improve, reduce, or have no effect on classification performance?
Explain your reasoning. Answer for both 2D and 3D for each of the three datasets.
2. k-Means
(a) Use k-Means as a feature transform, and visualize the transformed data in 2D and 3D for each of
the three datasets (like you did for PCA). Again, use the “with labels” version so that the class
label of each point is clear.
(b) Compare k-Means to PCA as a feature transform for each of the three datasets. Do you think
it would be better to use PCA or k-Means to transform the data into 2D and 3D? Specifically,
do you think it would be better in terms of classification performance on the transformed data?
Explain your reasoning, and answer for both 2D and 3D for each dataset.
(c) The primary use for k-Means is not as a feature transform, but as a clustering method. Use the
Normalized Mutual Information metric, described above, to get a quantitative measure of cluster
quality when using the same number of clusters as classes for each of the three datasets. Report
the performance, explain why you got the results you did.
3. Bayes Nets
(a) Consider the following four random variables: Cough, F ever, Allergies, Flu. You know that a
cough can be caused by either allergies or the flu, fever can only be caused by the flu, and
allergies and the flu have no causal relationship. Draw a Bayes Net that represents this set of
random variables and their independence/dependence relationships.
(b) Using the Bayes Net from the previous question, write down how you would factor the joint probability distribution, p(Cgh, F ev, Alg, Flu), as a product of conditional probabilities and marginals
given by the Bayes Net.
(c) In standard Hidden Markov Models, the three distributions you need to do any of the inference
tasks we talked about (filtering, smoothing, prediction, learning) are the transition model:
p(Zt = k | Zt−1 = j) = Ajk
the observation model:
p(Xt = l | Zt = j) = Bjl
and the prior :
p(Z0 = j) = πj
A simple modification of HMMs called Input/Output HMMs (or IOHMMs) changes the assumptions slightly. Instead of assuming transitions between hidden states are completely hidden, the
model assumes that transitions can be at least partially influenced by some observable random
variable, Ut. The Bayes Net for this model is given in Figure 2. Under the IOHMM model, what
distributions do you need? Describe them precisely in probability notation as was done above for
HMMs.
(d) In class we talked about how observing a common effect can make two marginally independent
random variables not conditionally independent (conditioned on the common effect). A simple
example of this is the XOR function on two bits. Let X1 and X2 be random variables that take
the values 1 or 0 with equal probability, and let Y be a random variable which is the XOR of X1
and X2.
Show that, given Y , the joint probability of X1 and X2 is not the same as the product
of X1 given Y and X2 given Y . That is, show that
p(X1, X2 | Y ) 6= p(X1 | Y ) · p(X2 | Y )
For reference, the Bayes Net for this problem is given in Figure 3.
Figure 2: Bayes Net for IOHMMs.
Figure 3: Bayes Net for the XOR question.