In 1996, a team of researchers began testing probabilistic models to sort incoming email messages. They relied on a core idea that would define the field for decades: naive Bayes classifiers assume every feature in a dataset is independent of all others once the target class is known.
Engineers apply logarithms to transform products into sums to solve this problem. This log-space computation preserves precision while avoiding rounding errors and allows rapid calculation even when dealing with thousands of potential features.
Multiple programs released in 1998 addressed growing email spam problems using naive Bayes techniques. Server-side solutions like DSPAM, Rspamd, and SpamAssassin incorporate similar probabilistic methods alongside modern mail client implementations.
Gaussian naive Bayes assumes continuous attributes follow a normal distribution defined by mean and variance parameters calculated from training sets. For example, height and weight measurements can be segmented by gender to compute specific statistical properties for each group.
Bayesian poisoning involves sending emails filled with legitimate text gathered from news sources or literature to decrease overall spam scores. Random innocuous words inserted into messages allow problematic content to slip past detection despite the presence of spam triggers.