This package contains Miroslaw's code base for Kaggle's see-click-fix competiton

This has been tested on ubuntu 12.04 LTS and should work on any standard linux distribution

Installation

sudo apt-get install python-scipy ipython python-sklearn python-pandas
git clone git@github.com:beegieb/kaggle_see_click_fix.git

Running the winning model

unzip -d data data/train.zip 
unzip -d data data/test.zip
unzip -d data data/geodata.zip
unzip -d data data/bryans_data.zip
mkdir cache, submissions

python test.py ridge_39

This will create all datasets required for the model, train the model, and save the predictions. The final model and intermediate datasets will be saved to cache.

By default the predictions will be saved to submissions/ridge_39.csv In general predictions are saved using the format $SUBMITDIR/$MODELNAME.csv

To change the default location for predictions and cache modify the global variables in config.py

I currently do not have any cache management implemented, so stale cache files are not removed. If you would like to try running a model with any other dataset configurations you will need to manually flush the cache by removing all relevent cache files in cache/. A safe bet is to delete everything in cache if you're unsure what is getting modified.

A note on the winning model: It was recently discovered that when I started using Bryan's dataset to explore his population-based features with my model there were some preprocessing steps he took with the summary and description fields I did not notice. As a result, to get exactly reproducable results for the winning model, you have to run the code with Bryan's data. There is a boolean flag inside of config.py that controls this. If you would like to try out my model without relying on Bryan's data you can set the boolean to False.

Accessing the model and data in python

import models
model = models.train_model('ridge39')
# I know the naming is poor at the moment, but after the model has been 
# trained once it will be pulled from cache

# The same goes for datasets associated with the model, they get pulled
# from cache after being created once
train_data, test_data = models.get_model_data('ridge39')

A brief summary of the code

All dataset and model configurations are set in settings.json

settings.json is a json object with two nested objects "models" and "datasets"

The format for defining a dataset is

"DatasetName": { "input_data": ["Names", "Of", "Input", "Datasets"]
                 "transforms": [["transform_name1", {"args": "values"}],
                                ["transform_name1", {"args": "values"}]] 
               }

The format for defining a model is

"ModelName": { "model": "PythonNameForModelObject", 
               "dataset": "NameOfDatasetUsedByModel",
               "target": "NameOFDatasetForTargetVariables",
               "args": {"NameOfArgument": "ValueOfArgument", ...},
               "validator": { "name": "NameOfCrossValidaton",
                              "args": "ArgsForCrossValidation" }
             }

Then these datasets and models can be used within python as shown above