kcrum/my_sshforest

Attempts at a Kaggle Problem

Kaggle is hosting a competition to predict which type tree---one out of seven possibilities---one would expect to encounter on 565,893 plots of Roosevelt National Forest in Colorado, where each plot measures 30 meters x 30 meters. A training set of 15,121 classified plots was provided. Some of the predictive variables, such as the elevation of the plot, are quantitative. Others variables are qualitative, such as which type of soil out of forty possibilities is present at the plot.

This repository contains some algorithms which attempt to solve this classification problem. All algorthims depend on the NumPy/SciPy/Matplotlib framework; most use scikit-learn. Most algorithms have a main() function that is invoked when the user calls

python somealgo.py

from the command line (here "somealgo.py" corresponds to one of the classifiers in this repository). The following pieces of code contain a classifier that can be called this way:

• "naive_bayes.py"
• "nearest_neighbor.py"
• "simple_forest.py"
• "simple_svm.py"

The best performing of these algorithms is "simple_forest.py," which is an random forest pulled straight out of the box from scikit-learn.

nielsen_net.py

The neural net in "nielsen_net.py" depends on sample code by Michael Nielsen, shared here on github. Currently you can't simply fork this repository and call "nielsen_net.py" without also setting up Nielsen's code on your machine.

sshforest_utilities.py

This contains a few useful plotting and munging functions. Calling

python sshforest_utilities.py

will make a cool pair of plots. One plot shows the contours of a distance-weighted 4 nearest neighbors classifier on 'Elevation' vs 'Horizontal_Distance_To_Hydrology' axes. The adjacent plot shows the distribution of training data over these same axes.