Probabilistic Models
Models that describe data with probability distributions, so they can say how likely each answer is and reason about uncertainty.
Also probabilistic graphical models · Bayesian models · generative classifiers
A weather forecaster who doesn't just say 'rain', but '70% chance of rain', because they keep track of how likely each outcome is.
Where it fits
+All relationships as a list
Is a kind of / part of
Kinds and parts
Siblings
Uses / built on
How it works
A probabilistic model says not just what the answer is, but how likely each possible answer is. It writes down a story of how the data could have been generated, in the language of probability, and fits that story to the data.
Members in this atlas:
- Naive Bayes: assumes features are independent given the class. Crude, fast, surprisingly good for text.
- Gaussian mixture models: the data comes from a few bell-shaped clusters.
- Hidden Markov models: a hidden state changes over time and emits what you observe.
Because many of them model how the data itself is distributed, they are generative: they can produce new samples, not just labels.
Everyday use cases
- Case 01
Spam filtering
Naive Bayes weighs the probability of each word appearing in spam versus normal mail.
- Case 02
Speech and sequences
Hidden Markov models infer the hidden words or states behind a sequence of sounds or measurements.
Strengths, weaknesses and when to use it
Strengths
- Output calibrated-ish probabilities and handle uncertainty and missing data naturally.
- Often work well with little data thanks to strong assumptions.
Weaknesses
- Those same assumptions (independence, Gaussian shapes, Markov chains) can be badly wrong.
- Usually less accurate than modern discriminative models on large datasets.
When to use itSmall datasets, problems where uncertainty matters, or when you need a model of how the data is generated.