docs/preprocessing.md

In jlmelville/sneer: Stochastic Neighbor Embedding Experiments in R

title: "Preprocessing" output: html_document: theme: cosmo toc: true toc_float: collapsed: false

Previous: Data Sets. Next: Input Initialization. Up: Index.

Sneer doesn't provide that much in the way of options for preprocessing, but it's useful to have some of this exposed because it can affect the final look of the embedding and you may need to scale data in a certain way if you're comparing with other embedding methods.

The following options mainly apply to input supplied as a data frame. If you supply a distance matrix, sneer won't do anything to it.

Duplicate Observations

First, some preprocessing you don't need to to do: some embedding methods ask you to weed out duplicates in the data before proceeding, but sneer tries to be robust enough to avoid that.

Low Variance Column Filtering

However, if you have some columns with very low or zero variance (for instance, because one column contains identical values for all observations), that can cause some problems. Rather than require you to remove these ahead of time, these will be removed from the data frame automatically before other pre-processing takes place. Sneer will log to the console the number of columns it filters in this way.

The scale_type Parameter

By default, no scaling is applied to the input data. By providing an argument to scale_type you can scale the data in the following way:

• "none" No scaling.
• "sd" Each column is centered so that is has a mean of zero and then scaled so that its variance is 1. I've seen this called "autoscaling" in chemical and biological data analysis, but I've not seen it referred to that way in other fields. Nonetheless, you may also refer to it by "auto" as well.
• "range" Range scales each column of data: each column is scaled so that the values in each column have a range from 0 to 1.
• "matrix" Range scales the entire matrix: like range scaling, except the entire matrix is treated like one big column.
• "tsne" Columns are centered and then each element divided by the absolute maximum value, so that all elements range from a bit less than -1 to 1, or -1 to a bit less than 1, depending on whether the maximum value was negative or positive. This is the scaling carried out in Barnes-Hut t-SNE.

All these arguments can be abbreviated (e.g. to "m" instead of "matrix").

Image data is often treated by range scaling the entire matrix, e.g. using the mnist data set as an example:

res_mnist <- sneer(mnist, scale_type = "m")

What sneer Doesn't Do

Some data sets can be very high dimensional, and processing time can be greatly reduced by doing PCA on the input data and keeping only a certain number of the score vectors that result.

Similarly, it's common to carry out various forms of whitening on image data sets.

While this could become part of the preprocessing workflow, I've kept this out of sneer for a couple of reasons.

First, if you run an embedding with different parameters on the same dataset more than once, it's a bit of a waste of time to repeat PCA and/or whitening inside the internals of sneer rather than have you do it yourself once before doing any further embedding.

Second, there's often a subjective criteria involved in this form of preprocessing. Everyone has their own favorite way of doing it, so rather than pointlessly expand the options even further, you should once again just do it yourself once ahead of time.

At least that was a nice and gentle start to the option-wrangling. Having preprocessed the input data, we can now move onto initializing the input data that the embedding directly works on.

Previous: Data Sets. Next: Input Initialization. Up: Index.

jlmelville/sneer documentation built on Nov. 15, 2022, 8:13 a.m.