Code of the winning entry to the Animal Behavior Challenge (ABC2018) for understanding animal behavior. Our approach is described in our paper.

The dependencies can be installed with:

pip install -r requirements.txt

Here is a brief description of the Python files in the archive:

• readme.md: this file.
• requirements.txt contains list of packages need to install to run all models.
• data/ contains original training and testing dataset along with newly created dataset.
• doc/ contains the technical report of the winning solution.
• output/ contains intermidiate outputs from trained model to use in the final submission.
• Submission/ contains the submission file to score.
• feature_engineering_day_night_split.py: creates new dataset with feature engineering by splitting original dataset to day and night. This will create a new dataset.
• feature_engineering_day_night_together.py : creates new dataset with feature engineering considering the original dataset without day and night split. This will create a new dataset.
• stacking_model_*.py : scripts for various modeling on the datasets
  • stacking_model_cat.py : train a CatBoost model for binary classification
  • stacking_model_lgb_gbt.py : train a LightGBM model with traditional Gradient Boosting Decision Tree for for binary classification
  • stacking_model_lgb_rf.py : train a LightGBM model with Random Forest for for binary classification
  • stacking_model_xgb_logistic.py : train a XGBoost model with logistic regression for binary classification
  • stacking_model_xgb_rank.py : train a XGBoost model to do ranking task by minimizing the pairwise loss
  • stacking_model_sk_gbt : train a GradientBoostingClassifier model from scikit-learn to do binary classification
  • stacking_model_sk_rf.py : train a RandomForestClassifier model from scikit-learn to do binary classification
  • stacking_model_sk_et.py : train a ExtraTreesClassifier model from scikit-learn to do binary classification
  • stacking_model_sk_svc.py : train a Support Vector Machines model from scikit-learn to do binary classification
  • stacking_model_sk_gpc.py : train a Gaussian process classification model from scikit-learn to do binary classification
• training.py : calls stacking_model_*.py scripts for two datset by passing parameter ("split" and "together")
• split_fn_param.py contains the optimal parameters from 5 fold cross validation, used in the stacking_model_*.py
• together_fn_param.py contains the optimal parameters from 5 fold cross validation, used in the stacking_model_*.py
• generate_submission.py generates final submission file from ensemble of files

Run following scripts in sequence:

1. Run feature_engineering_day_night_split.py to generate the csv files train_df_day_night_split.csv and test_df_day_night_split.csv.
2. Run feature_engineering_day_night_together.py to generate the csv files train_df_day_night_together.csv and train_df_day_night_together.csv.
3. Run training.py to train all models on datasets created from previous setps.
4. Run generate_submission.py to generate the final submission file.