🛫 ATOW_ML 🛬

Team repo for ATOW ML challenge (team_zealous_watermelon).

This work has been conducted for the OpenSkyNetwork PRC Data challenge, to try and predict ATOW the best way possible. This work is being made available to anyone under the GNU General Public License v3, "giving legal permission to copy, distribute and/or modify it", as long as these freedoms are passed on.

Open Sky Network: https://opensky-network.org.

Quick start

main.py contains all the code to run the models. We have built a pipeline that pre-processes and encodes the data, and then tunes the model parameters (XGBoost) to get the best possible results.

Another way to run models is via the jupyter notebooks that take the same form as main.py.

Requirements

⚠️ : Please make sure to put challenge_set.csv and submission_set.csv in the data folder of the ATOW_ML directory.

You need to have Python installed, with a version > 3.10.

To install the required packages, please run :

pip install -r requirements.txt

Data preprocessing & encoding

Preprocessing

We have added several adjustment to the original given dataset in order to prepare the data for training with XGBoost:

1. Converted timezone to local time and adding features such as day of the year, week, month to account for eventual seasonality in add_localtime_to_train_and_test
2. Computed longitude and latitude for each airport in compute_lon_lat
3. Simplified country_codes by group and rename countries in group_and_rename_countries
4. Simplify airport codes by group and rename airport in group_and_rename_airports
5. Regrouped less used airlines and aircraft types and create "XXXX" category for unknown ones in group_and_rename_aircraft_types

Encoding

We have chsoen to use hashing technique in order to convert data into fixed-size numerical values using a hash function (see string_to_int_hasing function for details).

Algorithm choice and model selection

We have chosen to use the model XGboost to predict the TOW. Model parameters can be found in models/xgboost_aggregation.py.

To make the predictions, we have specified and trained three different XGboost models :

• One on the whole dataset (xgboost_agreggation_model_2.json)
• Two others on H or M wtc
  • xgboost_agreggation_model_0.json is trained on wtc = H,
  • xgboost_agreggation_model_1.json is trained on wtc = M,

Bayesian optimization is applied for hyperparameter tuning to each model, using the Hyperopt module (see tuning.ipynb file in the notebooks folder).

We then combine them linearly to minimize RMSE. The weights are optimized for best performance (the optimal weights turn out to be 1/2 and 1/2).

Feel free to reach out to us for any inquiry!