docs/exported-data.md

In jlmelville/sneer: Stochastic Neighbor Embedding Experiments in R

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

Previous: Reporting. Next:Analysis. Up: Index.

sneer return value

Having run an embedding:

iris_tsne <- sneer(iris)

what exactly does iris_tsne give you? It's a list with something like the following:

• coords - a matrix of all-important coordinates. It's dimensions are n x ndim where n is the number of observations that were embedded and ndim is the number of output dimensions. That will normally be 2.
• method - a string containing the embedding method name.
• cost - the final cost associated with the output configuration in coords. See the Reporting section for more on this.
• norm_cost - the normalized version of cost. Again, see the Reporting section for more.
• iter - the iteration number at which the embedding stopped.
• nf - the number of cost function evaluations.
• ng - the number of cost gradient evaluations.

nf and ng only count the function and gradient values with respect to the error in the coordinates. Extra function and gradient evaluations made as part of optimizing "dynamic" parameters, if they were optimized separately from the coordinates (i.e. you set alt_opt = TRUE), are not included here. See the Embedding Methods section for details on which methods are (or can be made) dynamic in this way. See the section on the dyn return value below to get these separate function and gradient counts.

ret

You can also ask for some other data. Some of it might be useful for diagnostics, debugging, visualization or as input to another algorithm. To get access to it, pass a vector of names to the ret parameter.

The names you can ask for are:

• pcost - The final cost, decomposed into the sum of n values, with n being the number of points. Be aware that these individual components don't have to be positive, e.g. if the cost function is a divergence. For the pca method, the mmds cost function is used.
• x - input coordinates after Preprocessing and column filtering.
• dx - input distance matrix. Calculated if not present. Note that this of class matrix, not dist.
• dy - output distance matrix. Calculated if not present. Note that this of class matrix, not dist.
• p - input probability matrix. Only available if using a probability-based embedding.
• q - output probability matrix. Only available if using a probability-based embedding.
• prec - vector of input kernel precisions. Corresponds to the values summarized as prec in the console log after input initialization. Only available if using a probability-based embedding. See the Input Initialization section for more on the exact definition of the precision.
• dim - vector of intrinsic dimensionalities for each observation, calculated according to the method given in the multiscale JSE paper. These are meaningless if not using the default exponential \code{perp_kernel_fun}. See the Input Initialization section for more.
• deg Vector of degree centrality of the input observations as used in ws-SNE. Calculated if not present. A summary of this vector appears in the console log after input initialization if carrying out ws-SNE (by setting method = "wssne").
• dyn Dynamically optimized non-coordinate parameters. A list containing sublists, with names corresponding to the names of the parameters optimized. This is available if using you used the dyn input variable to make an embedding method "dynamic", or if you selected Inhomogeneous t-SNE (method = "itsne"), or DHSSNE (method = "dhssne"). For it-SNE, dyn will contain a list of dof, which contains the optimized degrees of freedom parameter for each output coordinate. If using method = "dhssne" method, dyn contains a list called alpha, which contains the optimized global alpha value. See the Embedding Methods for more. If the parameters were optimized separately from the coordinates, then this list will also contain the number of function and gradient evaluations used for parameter optimization under nf and ng, respectively.
• costs All the costs which were logged to screen during the optimization, as a matrix, with the iteration number in the first column.

For every name you provide, the list returned by sneer will contain an extra item in the list with the same name. If you ask for an item that doesn't make sense, e.g. a probability matrix from an embedding method that never calculated one, the request is silently ignored. Otherwise, the value will be calculated if possible, e.g. even though degree centrality is only used in ws-SNE, you can ask for it to be returned from any probability-based embedding, because it's related to the input probabilities, so it can be calculated.

# return the point-wise cost function, output distance matrix and the degree centrality
# in tsne_iris$dy and tsne_iris$deg, respectively
tsne_iris <- sneer(iris, ret = c("pcost", "dy", "deg"))

# returns the output distance matrix in sammon_iris$dy, but no
# sammon_iris$deg, because that would require the presence of data that
# Sammon Mapping doesn't create
sammon_iris <- sneer(iris, ret = c("dy", "deg"))

Although sneer requires input distance matrix data to be of class dist, internally, it uses objects of class matrix for all matrices, even distance matrices. So dy and dx will be of class matrix. If you want to pass dy or dx back into sneer for some crazy scheme or other, you will need to cast them into dist objects via as.dist:

s1k_tsne <- sneer(s1k, ret = c("dy"))
# ... work some unseen magic on s1k_tsne$dy ...
# return it to sneer for more processing
s1k_mmds <- sneer(as.dist(s1k_tsne$dy), method = "mmds", labels = s1k$Label)

You may find some use for the numerical vectors that can be returned with ret by coloring the embedding plot based on their magnitude. The Visualization section can tell you more.

Previous: Reporting. Next:Analysis. Up: Index.

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