RARE EVENT DETECTION:

This repository uses two methods to make a rare-event classifier. Both of them use reconstruction losses from the input data.

The dataset used in the repository can be downloaded from the link given here. https://docs.google.com/forms/d/e/1FAIpQLSdyUk3lfDl7I5KYK_pw285LCApc-_RcoC0Tf9cnDnZ_TWzPAw/viewform

Courtesy of --- Dataset: Rare Event Classification in Multivariate Time Series - Chitta Ranjan, Mahendranath Reddy, Markku Mustonen, Kamran Paynabar, Karim Pourak https://arxiv.org/abs/1809.10717

1. The first model 'ae_rse' uses a autoencoder for reconstruction loss. The cases where the reconstrucion loss is very high is treated as anamoly(in our case a break in the system).

2. The second model uses LSTM layers to regenerate the models simpy without using the encoder layer. This gives us a better accuracy as compared to the last case.

Please refer: https://towardsdatascience.com/extreme-rare-event-classification-using-autoencoders-in-keras-a565b386f098 for better understanding.

The best roc-auc score of 0.78 is achieved using specific threshold value.

The confusion matrix for the same is shown below.

How to use the model.

Training

On cloning the repo the following directory structure can be seen.

To train on your model delete the output_data folders and run the train.py file. You can change the model architecture in the model_utils.py file.

Run train.py to which will generate the model for both modes ae_rse and rse modes in separate directories. Or you can also download the model from the google drive link provided here which contains the entire output folders for both the case.

Testing

To test the model run the file - predict.py it will create a confusion matrix of the labels. Note you must have the data renamed as rare_events.csv inside the data folder to run the predict.py file.

Results

Below are my results: I achieved a roc_auc_score of 0.787 from the rse model and a roc_auc_score of 0.776 with the threshold values of 0.126 and 0.11 respectively.

The confusion matrix for the same are shown below:

1. AE_RSE model

Tensorboard logs:

confusion matrix:

1. RSE model

Tensorboard logs:

confusion matrix:

Future Goals:

Making an ensembele of both the models and generate even better results. Moreover the threshold value for each model can be further tuned for better results.