Gaussian Mixture Model for Model Monitoring

The Gaussian Mixture Model (GMM) is used for the Model Monitoring methodology.

A GMM is a trained probability density function of the input data seen during model training. It fits a mixture of Gaussian distributions to the training data and represents the operating domain covered by this data.

During inference, the trained GMM can be used as an indicator of the trustworthiness of the main model prediction. A very low probability density value indicates with high certainty that the current input state is outside the domain covered by the training data.

The output of the GMM is shown as log-probability. A threshold can be defined for this log-probability. If the output falls below the threshold, an anomaly is detected.

The continuous signal can also be converted into a binary signal. In this case, the value 1 indicates an anomaly.