Poisson Naive Bayes
When your input features are counts—like “how many times did this word appear?” or “how many purchases happened this week?”—it’s natural to model them as events happening a certain number of times. Poisson Naive Bayes is a Naive Bayes classifier built exactly for that kind of data.
What it assumes and how it works
Like all Naive Bayes models, it uses Bayes’ rule to pick the class with the highest posterior probability, and it makes a strong simplifying assumption: features are conditionally independent given the class. The “Poisson” part means each feature count x_j is modeled as a Poisson distribution whose rate depends on the class: λc,j. Training is straightforward: for each class and feature, estimate λc,j from the average count observed in that class (usually with smoothing so unseen events don’t force zero probability). At prediction time, it multiplies (adds log) Poisson likelihoods across features and combines them with the class prior.
Where you see it in practice
- Spam detection: counts of words, links, or exclamation marks per email.
- Fraud detection: number of transactions, chargebacks, or login attempts in a window.
- Customer churn: counts of support tickets, app opens, or failed payments.
Why it matters (and what can go wrong)
Choosing a distribution that matches your feature type matters because it shapes the likelihood calculation. If your counts are highly overdispersed (variance far larger than the mean) or features are strongly correlated, Poisson Naive Bayes can become overconfident and mis-rank classes. It’s a fast, interpretable baseline for count data, and a useful alternative to Multinomial Naive Bayes when you want per-feature Poisson rates rather than a “bag-of-words totals to N” assumption.
Poisson Naive Bayes is a Naive Bayes classifier that models each feature as a conditionally independent Poisson-distributed count given the class, estimating a rate parameter (λ) per feature and class. It is suited to sparse, nonnegative integer data such as event or token counts. It matters because matching the feature distribution to count data yields better-calibrated likelihoods and more accurate class posteriors than Gaussian-based assumptions.
Imagine sorting emails by counting things: how many times certain words appear, how many links they contain, or how many exclamation marks show up. Poisson Naive Bayes is an AI method that does this kind of sorting when the clues are simple counts.
It’s used for classification tasks—picking a label like “spam” vs “not spam” or “fraud” vs “normal.” The “Poisson” part means it treats each count like the kind of number you get from events happening over time or space (like calls per hour). The “naive” part means it assumes each clue acts independently, which often works surprisingly well in practice.