Anomaly Detection

Analytics Zoo shows how to detect anomalies in time series data based on RNN network. Currently, a Python notebook is provided. In the example, a RNN network using Analytics Zoo Keras-Style API is built, and NYC taxi passengers dataset is used to train and test the model.

We split the entire curve into 2 sections - training data and testing data. The training data section is treated all as normal and an RNN model is trained to fit the training data, the model has three LSTM layers followed by one Dense layer at the end.

model = Sequential()
model.add(LSTM(input_shape=(input_dim1, input_dim2, output_dim=8, return_sequences=True))

Train the model using using MSE loss (as a regression problem).

model.compile(loss='mse', optimizer='rmsprop'),y_train,batch_size=3028,nb_epoch=30)

Use this RNN model to predict the testing curve.

predictions = model.predict(x_test)

Anomalies could be defined by comparing the predictions and actual values. The current example defines data points as anomalies if the difference of predictions and actual values are larger than a certain value(threshold). See more details in the exmaple Python notebook