oracle-audience

This CLI tool odc and python script is meant to be used for the Oracle Audience Competition. Below are usage instructions

Installation

This script has been tested with Python 3.5

$ git clone git@github.com:CoDataScience/oracle-audience.git
$ python setup.py install

Usage

$ odc --help
Usage: odc [OPTIONS] COMMAND [ARGS]...

Options:
  --help  Show this message and exit.

Commands:
  sample
  score

Scoring

Using files in the correct input format and using the spend files provided in slack

$ odc score --help
Usage: odc score [OPTIONS] SPEND_FILE SUBMISSION_FILE

Options:
  --ratio  Score by computing top K using ratio instead of K=100,000
  --help   Show this message and exit.

$ odc score data/all_val_spend.csv vw_data/submission.csv

Below is sample output

Revenue: 10184379.730004184
Fraction of Possible Revenue: 0.6169140659646304
Number of Responders: 57033
Fraction of Possible Responders 0.5721379559407729

If you are using sampled data then the first file should only include what you want to score against and you should use the --ratio flag. Otherwise its expected you give exactly 100,000 advertisements

Sampling Data

For convenience there is also a script to sample data so you can model on smaller datasets that have the correct class balance. This script was used to generate the sampled data provided and can be used as below

$ odc sample --help
Usage: odc sample [OPTIONS] N_SAMPLES INPUT_PATH OUTPUT_PATH

Options:
  --seed INTEGER
  --help          Show this message and exit.

$ odc sample 10000 data/train data/sampled_train.txt

This will sample 10000 examples from data files that are in data/train. The files in there can be decompressed or still compressed in bz2 format. The script handles both.