Drift detection treated as a neural network regression problem

Winter 2020, semster project at ICT4SM Lab, EPFL

Neural network methods for anomaly detection in time series. In this work, 3 neural nets methods were explored:

• Boostrap Method
• Lower-Upper bound estimation (LUBE)
• LSTM Reconstruction error method

Please refer to the report for more information.

The implementation for the best resulting method (Bootstrap) was made public in this repo.

Requirements:

• Python 3.x
• Pytorch
• Pandas
• Numpy
• Matplotlib