Hokuspokus

Feature transformation for supervised learning using equation discovery.

For motivation, background and results see the paper or the (much) longer thesis.

Setup

You need a Common Lisp compiler, like SBCL. To install SBCL with homebrew:

brew install sbcl

Create a link in the default ASDF systems definition folder to the hokuspokus module:

ln -s ~/projects/hokuspokus/ ~/common-lisp/

Run

Start SBCL and load the hokuspokus module:

sbcl

(require :asdf)
(asdf:load-system :hokuspokus)

Example run using the kepler` dataset:

(hokuspokus:with
   (hokuspokus::in-file #p"data/kepler.arff")
   (hokuspokus::operators '+ '- '* '/ 'sqrt)
   (hokuspokus::pre-select :percent 0.25)
   (hokuspokus::post-select :abs 4)
   (hokuspokus::depth 3))

Look at the resulting feature space:

hokuspokus::*feature-space*

One of the resulting formulas can be simplified to Kepler's third law.

Kepler's Third Law: r^3/T^2

Matching formula:

(((r * r) * (r * r)) / ((t * r) * t))

Reduces to:

(r * r * r) / (t * t)

Weka integration

Data files are read and written in the Weka ARFF format.

To write the results of a run back to an ARFF file:

(hokuspokus:with
   (hokuspokus::in-file #p"data/kepler.arff")
   (hokuspokus::operators '+ '- '* '/ 'sqrt)
   (hokuspokus::pre-select :percent 0.25)
   (hokuspokus::post-select :abs 4)
   (hokuspokus::out-file #p"data/kepler.processed.arff")
   (hokuspokus::depth 3))

Limitations

Only numerical, non-sparse data sets are supported.