ML methods for prediction, using categorical variables and time

Question

Most of the time series analysis tutorials/textbooks I found time series data, usually deal with continuous numerical variables. I am currently trying to solve a problem that deals with multivariate time series data, where the fields are all categorical variables.

Specifically, my data is a stream of alert data, where at each time stamp, information such as the alert monitoring system, the type of alert, the location of the problem, etc. are stored in the alert. These fields are all categorical variables.

My data is very similar to the one described in another question, so I used a similar description. However my question is a bit more general.

Given a chain of alerts, what is a good approach for using ML to predict the location and type of the next alert (at t(1) ), having knowledge of common alert patterns and the values for t(0), t(-1), t(-2), ...t(-n) ?

Answer 1

One way to frame the problem of predicting patterns of complex discrete states is probabilistic graphical models (PGMs). PGMs model joint (multivariate) distributions over large numbers of random variables that interact with each other.

PGM do not have a notion of linear time. They model state. Given the current state and previous states, how likely are other states in the next time step? In your problems, state is the type and location of alert.

Common examples of PGMs are Hidden Markov Models (HMMs) and Conditional Random Fields (CRFs).