Multivariate normal distribution vs Gaussian Mixture Models

Hi there,

I think the lecture gave a great introduction to anomaly detection using the multivariate normal distribution. I was browsing online for more materials to dive into topic, and found one technique called Gaussian Mixture Model, which also seems to be a method for anomaly detection based on multivariate Gaussian distribution. I wondered if this is essentially the same thing as what was introduced in the lecture? If not, can anyone help me understand the difference, or point me to the resources to dive into these two topics? Thank you!

Hello @yyaolulu,

Let’s say we have 3 features, and if we assume a multivariate Gaussian model, we will have one Gaussian distribution for each feature. In contrast, if we assume a multivariate Gaussian mixture model, we have one or more Gaussian distributions per feature.

One example of using a Gaussian Mixture Model is to model the age distribution of people inside the elderly care industry, because we have two different populations - the service providers who are younger working force, and the service users who are elderly. In this case, we should see two peaks in the age distribution, so using a single Gaussian distribution would not be appropriate.