We have seen how principal component analysis (PCA) can be used in the dimensionality reduction task—reducing the number of features of a dataset while maintaining the essential relationships between the points. While PCA is flexible, fast, and easily interpretable, it does not perform so well when there are nonlinear relationships within the data; we will see some examples of these below.
To address this deficiency, we can turn to a class of methods known as manifold learning—a class of unsupervised estimators that seeks to describe datasets as low-dimensional manifolds embedded in high-dimensional spaces. When you think of a manifold, I'd suggest imagining a sheet of paper: this is a two-dimensional object that lives in our familiar three-dimensional world, and can be bent or rolled in that two dimensions. In the parlance of manifold learning, we can think of this sheet as a two-dimensional manifold embedded in three-dimensional space.
Rotating, re-orienting, or stretching the piece of paper in three-dimensional space doesn't change the flat geometry of the paper: such operations are akin to linear embeddings. If you bend, curl, or crumple the paper, it is still a two-dimensional manifold, but the embedding into the three-dimensional space is no longer linear. Manifold learning algorithms would seek to learn about the fundamental two-dimensional nature of the paper, even as it is contorted to fill the three-dimensional space.
Here we will demonstrate a number of manifold methods, going most deeply into a couple techniques: multidimensional scaling (MDS), locally linear embedding (LLE), and isometric mapping (IsoMap).
We begin with the standard imports:
To make these concepts more clear, let's start by generating some two-dimensional data that we can use to define a manifold. Here is a function that will create data in the shape of the word "HELLO":
Let's call the function and visualize the resulting data:
The output is two dimensional, and consists of points drawn in the shape of the word, "HELLO". This data form will help us to see visually what these algorithms are doing.
Looking at data like this, we can see that the particular choice of x and y values of the dataset are not the most fundamental description of the data: we can scale, shrink, or rotate the data, and the "HELLO" will still be apparent. For example, if we use a rotation matrix to rotate the data, the x and y values change, but the data is still fundamentally the same:
This tells us that the x and y values are not necessarily fundamental to the relationships in the data.
What is fundamental, in this case, is the distance between each point and the other points in the dataset.
A common way to represent this is to use a distance matrix: for points, we construct an array such that entry contains the distance between point and point .
Let's use Scikit-Learn's efficient
pairwise_distances function to do this for our original data:
As promised, for our N=1,000 points, we obtain a 1000×1000 matrix, which can be visualized as shown here:
If we similarly construct a distance matrix for our rotated and translated data, we see that it is the same:
This distance matrix gives us a representation of our data that is invariant to rotations and translations, but the visualization of the matrix above is not entirely intuitive. In the representation shown in this figure, we have lost any visible sign of the interesting structure in the data: the "HELLO" that we saw before.
Further, while computing this distance matrix from the (x, y) coordinates is straightforward, transforming the distances back into x and y coordinates is rather difficult.
This is exactly what the multidimensional scaling algorithm aims to do: given a distance matrix between points, it recovers a -dimensional coordinate representation of the data.
Let's see how it works for our distance matrix, using the
precomputed dissimilarity to specify that we are passing a distance matrix:
The MDS algorithm recovers one of the possible two-dimensional coordinate representations of our data, using only the distance matrix describing the relationship between the data points.
The usefulness of this becomes more apparent when we consider the fact that distance matrices can be computed from data in any dimension. So, for example, instead of simply rotating the data in the two-dimensional plane, we can project it into three dimensions using the following function (essentially a three-dimensional generalization of the rotation matrix used earlier):
Let's visualize these points to see what we're working with:
We can now ask the
MDS estimator to input this three-dimensional data, compute the distance matrix, and then determine the optimal two-dimensional embedding for this distance matrix.
The result recovers a representation of the original data:
This is essentially the goal of a manifold learning estimator: given high-dimensional embedded data, it seeks a low-dimensional representation of the data that preserves certain relationships within the data. In the case of MDS, the quantity preserved is the distance between every pair of points.
Our discussion thus far has considered linear embeddings, which essentially consist of rotations, translations, and scalings of data into higher-dimensional spaces. Where MDS breaks down is when the embedding is nonlinear—that is, when it goes beyond this simple set of operations. Consider the following embedding, which takes the input and contorts it into an "S" shape in three dimensions:
This is again three-dimensional data, but we can see that the embedding is much more complicated:
The fundamental relationships between the data points are still there, but this time the data has been transformed in a nonlinear way: it has been wrapped-up into the shape of an "S."
If we try a simple MDS algorithm on this data, it is not able to "unwrap" this nonlinear embedding, and we lose track of the fundamental relationships in the embedded manifold:
The best two-dimensional linear embeding does not unwrap the S-curve, but instead throws out the original y-axis.
How can we move forward here? Stepping back, we can see that the source of the problem is that MDS tries to preserve distances between faraway points when constructing the embedding. But what if we instead modified the algorithm such that it only preserves distances between nearby points? The resulting embedding would be closer to what we want.
Visually, we can think of it as illustrated in this figure:
Here each faint line represents a distance that should be preserved in the embedding. On the left is a representation of the model used by MDS: it tries to preserve the distances between each pair of points in the dataset. On the right is a representation of the model used by a manifold learning algorithm called locally linear embedding (LLE): rather than preserving all distances, it instead tries to preserve only the distances between neighboring points: in this case, the nearest 100 neighbors of each point.
Thinking about the left panel, we can see why MDS fails: there is no way to flatten this data while adequately preserving the length of every line drawn between the two points. For the right panel, on the other hand, things look a bit more optimistic. We could imagine unrolling the data in a way that keeps the lengths of the lines approximately the same. This is precisely what LLE does, through a global optimization of a cost function reflecting this logic.
LLE comes in a number of flavors; here we will use the modified LLE algorithm to recover the embedded two-dimensional manifold. In general, modified LLE does better than other flavors of the algorithm at recovering well-defined manifolds with very little distortion:
The result remains somewhat distorted compared to our original manifold, but captures the essential relationships in the data!
Though this story and motivation is compelling, in practice manifold learning techniques tend to be finicky enough that they are rarely used for anything more than simple qualitative visualization of high-dimensional data.
The following are some of the particular challenges of manifold learning, which all contrast poorly with PCA:
With all that on the table, the only clear advantage of manifold learning methods over PCA is their ability to preserve nonlinear relationships in the data; for that reason I tend to explore data with manifold methods only after first exploring them with PCA.
Scikit-Learn implements several common variants of manifold learning beyond Isomap and LLE: the Scikit-Learn documentation has a nice discussion and comparison of them. Based on my own experience, I would give the following recommendations:
If you're interested in getting a feel for how these work, I'd suggest running each of the methods on the data in this section.
One place manifold learning is often used is in understanding the relationship between high-dimensional data points. A common case of high-dimensional data is images: for example, a set of images with 1,000 pixels each can be thought of as a collection of points in 1,000 dimensions – the brightness of each pixel in each image defines the coordinate in that dimension.
Here let's apply Isomap on some faces data. We will use the Labeled Faces in the Wild dataset, which we previously saw in In-Depth: Support Vector Machines and In Depth: Principal Component Analysis. Running this command will download the data and cache it in your home directory for later use:
We have 2,370 images, each with 2,914 pixels. In other words, the images can be thought of as data points in a 2,914-dimensional space!
Let's quickly visualize several of these images to see what we're working with:
We would like to plot a low-dimensional embedding of the 2,914-dimensional data to learn the fundamental relationships between the images. One useful way to start is to compute a PCA, and examine the explained variance ratio, which will give us an idea of how many linear features are required to describe the data:
We see that for this data, nearly 100 components are required to preserve 90% of the variance: this tells us that the data is intrinsically very high dimensional—it can't be described linearly with just a few components.
When this is the case, nonlinear manifold embeddings like LLE and Isomap can be helpful. We can compute an Isomap embedding on these faces using the same pattern shown before:
The output is a two-dimensional projection of all the input images. To get a better idea of what the projection tells us, let's define a function that will output image thumbnails at the locations of the projections:
Calling this function now, we see the result:
The result is interesting: the first two Isomap dimensions seem to describe global image features: the overall darkness or lightness of the image from left to right, and the general orientation of the face from bottom to top. This gives us a nice visual indication of some of the fundamental features in our data.
We could then go on to classify this data (perhaps using manifold features as inputs to the classification algorithm) as we did in In-Depth: Support Vector Machines.
As another example of using manifold learning for visualization, let's take a look at the MNIST handwritten digits set. This data is similar to the digits we saw in In-Depth: Decision Trees and Random Forests, but with many more pixels per image. It can be downloaded from http://mldata.org/ with the Scikit-Learn utility:
This consists of 70,000 images, each with 784 pixels (i.e. the images are 28×28). As before, we can take a look at the first few images:
This gives us an idea of the variety of handwriting styles in the dataset.
Let's compute a manifold learning projection across the data. For speed here, we'll only use 1/30 of the data, which is about ~2000 points (because of the relatively poor scaling of manifold learning, I find that a few thousand samples is a good number to start with for relatively quick exploration before moving to a full calculation):
The resulting scatter plot shows some of the relationships between the data points, but is a bit crowded. We can gain more insight by looking at just a single number at a time: