Plan for v0.2.0

[lindahua]

This package is an important foundation for our efforts towards a machine learning ecosystem. I plan to spend some time to work on this package recently.

Here is a tentative plan:

• [ ] Remove dependence on DataFrames

  Particularly, this package should provide core dimensionality reduction algorithms that focus on ordinary arrays. Dependence on data frames should be removed.

• [ ] Consistent interface

  To conform with other machine learning packages, this package should use column-major data set format (i.e. each column being an observation).

• [ ] Improve PCA

  Implement multiple PCA algorithm (e.g. based on covariance, SVD, transposed data, etc), and a high-level pca function that selects an appropriate method based on input data.

• [ ] Linear Discriminant Analysis

• [ ] Independent Component Analysis

  It seems that a Fast ICA algorithm has been implemented. But will add testing codes.

• [ ] Canonical Correlation Analysis
• [ ] Factor Analysis & Probabilistic PCA

• [ ] Classic MDS

  It is already here. May need some testing.

• [ ] Separate NMF to another package.

  Nonnegative Matrix Factorization in itself is a big field that deserves its own package. I have created a package NMF.jl for this purpose, and implemented a more sophisticated framework there.

• [ ] Move manifold embedding functions to a separate package. These are a different family of methods.

These are the basics. I believe we can release v0.1.0 when these are ready.

cc: @johnmyleswhite

[johnmyleswhite]

Need to read this again, but all of it sounds reasonable on a quick read.

[lindahua]

I tagged the current version as v0.1.0, and new version developed under the new plan (which probably has breaking changes) will be tagged v0.2.0.

[ccsv]

I would say this would need additional features such as

Varimax rotation

http://en.wikipedia.org/wiki/Varimax_rotation

[lindahua]

Most of these have been implemented in MultivariateStats.jl