Overview / Usage

Symbolic Regression refers to the search for a mathematical expression that describes a nonlinear relationship between variables from some data set. The usual representation for a solution from the search space is the Expression Tree which is capable of representing every solution from the space of mathematical expressions. While the generalization power is desirable, a large portion of the search space is composed of very complicated expressions that are no better than a black box model. In this project we seek a restrictive representation that only allows simple expressions while keeping the space general enough to approximate the data relationship.

Methodology / Approach

The Interaction-Transformation representation leads to a rich representation of mathematical expressions capable of describing nonlinear relationships through a linear combinations of interactions of the original variables. This allows more expressiveness for the fitted model.
In this project different traditional regression algorithms will be adapted to search for IT expressions and all the IT-version and nonIT-versions will be compared with each other by means of Mean Absolute Error, Root of Mean Squared Error and Complexity of expression when applied to real world data sets.

Technologies Used

Python (3.6), Haskell (stack)

Repository

https://github.com/folivetti/ITEA