Chapter 6. Hidden Markov Models
Intuition informs much of what we do: for example, it tells us that certain words tend to be a certain part of speech, or that if a user visits a signup page, she has a higher probability of becoming a customer. But how would you build a model around intuition?
Hidden Markov models (HMMs) are well versed in finding a hidden state of a given system using observations and an assumption about how those states work. In this chapter, we will first talk about how to track user states given their actions, then explore more about what an HMM is, and finally build a part-of-speech tagger using the Brown Corpus. The part-of-speech tagger will tag words in sentences as nouns, pronouns, or any part of speech in the Brown Corpus.
HMMs can be either supervised or unsupervised and also are called Markovian due to their reliance on a Markov model. They work well where there doesn’t need to be a lot of historical information built into the model. They also work well for adding localized context to a classification. Unlike what we saw with Naive Bayesian Classification, which relied on a lot of history to determine whether a user is spammy or not, HMMs can be used to predict changes over time in a model.
Tracking User Behavior Using State Machines
Have you ever heard of the sales funnel? This is the idea that there are different levels of customer interaction. People will start as prospects and then transition into more engaged states (see Figure 6-1).