Rare Event Prediction

These models are adapted from these blog posts:

• Extreme Rare Event Classification using Autoencoders in Keras
• LSTM Autoencoder for Extreme Rare Event Classification in Keras

by Chitta Ranjan

The original source is included as Notebooks:

• autoencoder_classifier.ipynb
• lstm_autoencoder_classifier.ipynb

To train the models, use:

$ guild run ae:train
$ guild run lstm:train

The LSTM does not include validation accuracy.

To Do

• Generate sample log (treat as simulation problem)
  • Contains mostly normal log events of whatever (negative example)
  • Supports SIGTERM or some other signal
  • Prints signal
  • After some period with a random component, logs a "crash" (positive example)
• Convert simulated logs into format we can train
• Activation functions (elu, leaky relu, etc) (see advanced activations in Keras)
• More or fewer layers
• Different optimizers
• Within the LSTM:
  • Dropout
  • ???
• Bump epochs to 1000
• Add early stopping (Keras callback)
• Learning rate schedules
• Use custom Keras metic for roc_auc (unless slows training)
• Check if metrics for LSTM is slowing training

Bug in data processing

• Losing a column somehow
• He's using the row number in the xs, which masks the missing col

---

• Highlight feature engineering in data-preparation (convert from raw to prepared - time shift of y values)

• Use validation data for examples