Amazon Awesomeness Predictor

Description

This repository contains the code for data exploration & preprocessing, parameter tuning, and different approaches attempted to predict whether a product on Amazon is awesome or not (binary classification). Awesomeness is defined as whether the product's average review > 4.5 stars (in the available dataset). The final model (final_iteration_model.py) predicts using unseen training data (from full_train.csv) with an average weighted F-1 score of 0.76 (average result from 10-fold cross validation). On test data (from Test.csv) it predicts with a weighted F-1 score of 0.73.

The code in this repository is for a group project & personal extra credit submission for COSC 74 at Dartmouth College.

Instructions

1. Clone the repo.

2. cd amazon-awesomeness-predictor.

3. tar -xf data/Test.tar.gz and tar -xf data/full_train.tar.gz; this should extract full_train.csv and Test.csv into the same directory as main.py.

4. Run main.py with full_train.csv and Test.csv in the same directory to see the final model's performance (precision, recall, F-1) in each of the 10 folds. Change the third param in main.py from xgb10 to predict then rerun to produce the predictions CSV file for test data in Test.csv using the final model. The output file is formatted as follows: 1st column - automatic indexing by Pandas; 2nd column - Amazon product ID; 3rd column - Awesomeness.

If you would like to see the performance of the earlier models (first_iteration_model.py or second_iteration_model.py), simply run those files with full_train.csv in the same directory. These models were not adapted to produce a predictions CSV for Test.csv data because they were used not in the end.

Data Files

full_train.csv: training data for the models, contains data about Amazon product reviews e.g. review text, review summary, product price, etc. The binary target column indicates whether the product is awesome across all its reviews.

Test.csv: unlabeled test data.

Note that these have been compressed (separately due to upload size limits for individual files). The compressed files are in data/. Please follow the instructions above to extract them.

Code Files

final_iteration_model.py: code of the final model (contains both group project & personal extra credit code).

second_iteration_model.py: code of my group's second approach.

first_iteration_model.py: code of my group's first approach.

common.py: common code used by the different iterations above.

data_exploration.py: code of my group's data exploration, with comments explaining observations and conclusions.

data_preprocessing.py: preprocessing code, with comments explaining observations and conclusions. Most of this code is unused in the final model; most preprocessing done in the final model is directly included (with documentation) in final_iteration_model.py.