Predicting flight cancellations with a generic two-layer Artificial Neural Network

neural_network.py and activation_functions.py are generic. The artificial neural network is set to initialize to two layers, with 400 neurons in the first hidden layer and 250 neurons in the second hidden layer. This neural network architecture is configured to work with flight delays data, which load_data.py transforms into 368 binary input variables.

The neural network uses a leaky rectified linear unit activation function to compute the hidden layers, and a sigmoid activation function to compute the output layer (which is a binary classification in the case of predicting flight cancellations). It is straightforward to modify the specified activation function for the output layer (don't forget the associated derivative in back propagation), though care needs to be taken as the learning rate likely requires a lower initial value if the network is adapted to perform regression.

• To train the artificial neural network using k-fold cross-validation, execute train_network.py in command line (k = 2 is a single train/test split; k >= 2):

  python train_network.py 2
  

  The training data file './flight_delays_data.csv' is required. The user can specify a different file at line 112. The program's default behavior is to start a new training from scratch. To resume a training, modify line 115 accordingly and specify k at line 116 for the k-fold cross-validation in progress.

• To use a trained neural network to predict claims for delayed or cancelled flights, execute predict_claims.py in command line:

  python predict_claims.py
  

  The user will be prompted to specify the path to a .csv file with the test data. Weights for the trained network needs to be in ./state/best/.