latent variable
/ LAY-tunt VAIR-ee-uh-bul /
A latent variable is something you believe is influencing your data but cannot measure directly — it stays behind the curtain. You never observe 'customer satisfaction' or 'a person's underlying health' as a number; you observe their effects: reviews, repeat purchases, blood tests, symptoms. The latent variable is the hidden cause you posit to make sense of the visible clues, the unseen knob you assume is turning the dials you can read.
Why bother inventing something you can't see? Because it often explains a messy pile of observations far more simply. Suppose you measure dozens of physical and verbal test scores; rather than tracking all their tangled correlations, you might propose one or two hidden traits — say 'verbal ability' and 'spatial ability' — that, if known, would make the scores fall into place. The latent variables compress the chaos into a few meaningful dials, which is the core idea behind techniques from topic modeling to the compact 'codes' inside modern generative models.
Why it matters and where to be careful: latent variables are how probabilistic models capture structure that isn't spelled out in the raw data, and inferring them is exactly what algorithms like EM and variational inference are built to do. The honest caution is that a latent variable is a modeling choice, not a discovered fact. You can fit a model that posits 'intelligence' or 'personality types' and get clean numbers, but the existence of the hidden thing, and any story about what it 'really' is, are assumptions you imposed — not proof the curtain hides what you named.
Topic modeling of news articles. The words you can count are observed; the 'topic' of each article — sports, politics, finance — is a latent variable nobody labeled. The model assumes each article mixes a few hidden topics, and each topic favors certain words. Inferring the latent topics lets the system group articles it was never told the categories for.
A latent variable (the article's topic) is never observed directly but is inferred from the words it tends to produce.
A latent variable that fits the data well is not thereby proven real. Models that posit hidden 'factors' or 'types' can be useful and predictive while the named entity remains an interpretive assumption — never confuse a convenient hidden cause with a confirmed one.