See a detailed methodology at https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/discussion/52645.
~~
To reproduce the final solution, download data from https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/data and put it in data
folder, make a cache
and submit
folder. Download Neptune models from https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/discussion/51836. Zafar's cleaned data comes from https://www.kaggle.com/fizzbuzz/toxic-data-preprocessing/code. You also will have to download embedding files. (See more setup.txt
.)
Run pip install -r requirements.txt
, and then run all the models:
python lr.py
python fe_lgb.py
python sparse_lgb.py
python sparse_fe_lgb.py
python gru.py
python gru2.py
python gru80.py
python gru128.py
python gru128-2.py
python gru-conv.py
python double-gru.py
python 2dconv.py
python lstm-conv.py
python dpcnn.py
python rnncnn.py
python rnncnn2.py
python capsule_net.py
python attention-lstm.py
python fm.py
python ridge.py
python neptune-ml-models.py
python lvl2_final_cnn.py
python lvl2_final_lgb.py
After that, you will have a final submission ready to go. cache.py
and cv.py
maintain the model run layer and are imported, not called directly. feature_engineering.py
contains all the feature engineering, and is imported and run as needed. preprocess.py
and utils.py
contain additional functions that are used by the models.
lvl3_hillclimb_average.py
should be run after python lvl2_final_lgb.py
to find optimal weights. These weights are then manually normalized to sum up to 1 and then manually put back into python lvl2_final_lgb.py
.
relative_word_frequency_analysis.py
identifies words that are more frequently found in toxic than non-toxic comments. It was used to inform the word list found in feature_engineering.py
.
conversationai_api.py
contains our attempt to query the PerspectiveAPI for data. This ended up being unhelpful. To run this, you will need to set up API access.
lr-extra_data.py
contains our attempt to run linear regressions on the extra Wikipedia toxicity and agression labels. This ended up being unhelpful.