josh.meanings

Initialization

Initialization is implemented as a multimethod whose dispatch is chosen by the :init keyword. The following initialization methods are supported.

• :niave
• :k-means-++
• :k-means-parallel
• :k-mc-squared
• :afk-mc

If you would prefer a different initialization scheme, you can provide your own by adding a new defmethod.

When calling the library the implementation will default to afk-mc which is the assumption free approximation of k-means-++.

Dataset Format

This library assumes that datasets will be medium data - larger than memory, but smaller than disk. Callers must provide a reference to the file which contains the dataset or a lazy sequence so that the dataset is never fully realized in memory. The library leverages techascent.ml.dataset to handle dataset serialization and deserialization.

Some file types aren't well suited to high performance serialization and deserialization. The :format keyword controls what file format will be used to persist data during the computation. The list of supports formats are:

• :csv
• :arrow
• :arrows
• :parquet

By default the :parquet format will be used.

Distance Functions

Which distance function should be used is left up to a multimethod protocol that will dispatch based on the :distance-fn key. Special care should be taken when choosing non-euclidean distances, because k-means is not guaranteed to converge or stabilize with arbitrary distance functions.

The get-distance-fn multimethod dispatches based on identity to determine what distance function to use when calculating distances. By default the supported distance function keys are:

• :euclidean
• :manhattan
• :chebyshev
• :correlation
• :canberra
• :emd
• :euclidean-sq
• :discrete
• :cosine
• :angular
• :jensen-shannon

The default distance function is :emd which may not be appropriate for all use cases since it doesn't minimize the variational distance like euclidean would. If you don't know why :emd is the default you should probably switch to using :euclidean.

Calling Options

Though it is possible to run k means a single time this is not recommended. In general, k means doesn't find the globally optimal solution. it is especially prone to finding bad clusters when run using naive initialization, but even with k-means-++ and k-means-parallel there is stochasticity. Multiple runs can converge on different solutions. So it makes sense to run multiple times and keep the best one.

k-means-seq returns a lazy sequence of cluster results as well as their cost. You can use it in various ways to ensure you get a good clustering run - perhaps taking the first ten cluster attempts and keeping the best one. Or perhaps something more sophisticated like continuing to do clustering until you think the probability of finding an improvement goes below some threshold.

Testing

Tested on both massive datasets and toy datasets. This is an area that could use further improvement, since right now the tests are for the happy path - the default number of runs, init, and format.

To run unit tests use lein test.