Data Sets

Previous: Concepts. Next: Preprocessing. Up: Index.

Data Format

sneer can take two types of input data, a data frame or a distance matrix.

Data Frame

Each row should be an observation and each column a different variable. For examples of suitable data sets, see below. Sneer uses only numeric vectors for its calculations and ignores anything else. However, if you have a factor column or a column of colors (e.g. RGB strings), that data can be used to color the plot. See the Visualization section for more on that.

Distance Matrix

Alternatively, a distance matrix of class dist. For visualization you will want to come up with an appropriate vector from somewhere and pass it in via the colors or labels parameter. Again, see the Visualization section for more.

If the data is in the correct format, you are ready to sneer it. The next section covers Preprocessing. But below are some sources of potential datasets.

Sources of Data

Once you start experimenting with sneer, it’s useful to have multiple sources of data to try your ideas out on. The following are some data sets which researchers in the field have used or which I have known and loved.

iris

For just fiddling about, the good old iris dataset has the advantage of being small. But with only 4 numeric measurements per iris, that’s not a very high-dimensional data set. In low-dimensional situations, you’re unlikely to see the advantages of t-SNE. I suggest being very wary of drawing conclusions from data sets where something like PCA, metric MDS or Sammon mapping does a very good job. t-SNE just isn’t necessary under these circumstances. Lucky for you, sneer can carry out PCA, MDS and Sammon mapping while keeping all the other options the same, so you should be able to find out quite quickly if this is the case.

s1k

sneer comes with a data set called s1k. It’s entirely synthetic and designed to reproduce the “crowding problem” that t-SNE sets out to solve. The crowding problem is due to there being a lot more volume around a point in high dimensional space than in a 2D plot. Imagine a point in the middle of a cube with its nearest neighbors being on each of the six faces. There’s no way we can reproduce that in two dimensions, at best we can have four nearest neighbors, with the other two out in the comparative cold. The problem gets worse in higher dimensions.

s1k is a fuzzy 9D simplex: there are 10 points which are all equidistant from each other. Each of those ten points was then replicated another 99 times each, adding some symmetric gaussian noise to its coordinates. So we now have 1000 points, consisting of 10 clusters each. I made the spread of the gaussian sufficiently large so that the overlap between clusters gave traditional distance-based methods like PCA and metric MDS a hard time.

Without wanting to get too far ahead of ourselves, let’s have a look at how PCA, MDS and t-SNE get on with s1k:

s1k with PCA

Yuk. If you squint, you can sort of see something. There’s very noticeable crowding, with a lot of points crushed in the center of the plot. That’s characteristic of trying to stuff a high dimensional data set into 2D.

s1k with metric MDS

That’s a bit better. But if I hadn’t colored the clusters in the plot, you’d be hard pressed to tell me what structure the data had.

s1k with t-SNE

Yay. Much better. t-SNE really does make a difference.

So there you go. s1k passes my low bar for being a useful data set. We could (and probably should) make life a bit harder for ourselves. Real data will differ from s1k by having different numbers of points per cluster, different distances between clusters, different spreads of data in the cluster, and being anisotropic (the spread won’t be symmetric in all dimensions).

There are some other data sets out there you may find useful. In a fit of shameless self-promotion, I will mention some other R packages I wrote that help wrangle them into a sneer-able form.

None of the following packages are on CRAN, but are on github. To install them, install the devtools package:

install.packages("devtools")

Then use the devtools::install_github function to fetch them from github.

snedata

Find the snedata package at https://github.com/jlmelville/snedata and install it with:

devtools::install_github("jlmelville/snedata")

Under the snedata umbrella, there are three sources of data here you might be interested in:

JSE simulation data.

Lee and co-workers have published some papers on a technique similar to t-SNE called Jensen-Shannon Embedding - see the JSE and multi-JSE papers.

In these papers, they use some simple simulation data sampled from shapes, including a 3D ball, the surface of a sphere, and helical toroid (this looks like a spring that’s been stretched round in a circle so the ends meet). The snedata has data sets which are very similar to these.

“How to Use t-SNE Effectively”

A team at Google (Cloud and Brain) published an interactive article How to use t-SNE Effectively, containing some tips on the effect of parameters on t-SNE embeddings. You can run t-SNE directly in the browser on that page on a set of simulation data, much of it based around various gaussian clusters. I converted JavaScript code which generated these datasets, also available on github at https://github.com/distillpub/post--misread-tsne, into R, so these are also available for you to use.

Olivetti and Frey Faces

These are two sets of face data (images of different people’s faces taken at different angles) that have been used in the literature. They are available in the R package RnavGraphImageData but not in a format that can be directly used by sneer. The snedata package also provides functions to transform those datasets so sneer can use them. If you want to do this, you’ll need to install RnavGraphImageData manually via install.packages("RnavGraphImageData")

mnist

Find the mnist package at https://github.com/jlmelville/mnist and install it with:

devtools::install_github("jlmelville/mnist")

The MNIST dataset is a set of images of hand written digits. It (or a subset) is probably the most commonly used dataset used in the literature.

This package doesn’t distribute the MNIST dataset directly, but downloads it from its website and then formats it, so you’ll need internet access to get hold of it. Like the Frey and Olivetti data sets, there are also functions to view each digit.

coil20

Find the coil20 package at https://github.com/jlmelville/coil20 and install it with:

devtools::install_github("jlmelville/coil20")

Another image dataset (you may be detecting a pattern here), this time of 3D objects. Once again, this dataset doesn’t distribute the images, just downloads them and formats them for you. And it provides a function for viewing the images.

Between all these data sets, that should be enough data to be getting on with. Watch out for the large data sets - due to sneer’s design and pure-R nature, it’s easy to have it sit there churning away for hours, or just run out of memory. I suggest experimenting with sub sets of 100-1000 data points before committing to embedding the entire MNIST data set (hint: you probably don’t want to do that).

Previous: Concepts. Next: Preprocessing. Up: Index.