Docs / Obsidian Suite / For administrators
Learning
Obsidian Suite's Bayesian filter learns what your spam and your legitimate mail look like: words, link domains, senders, attachment types and mailer software. The more you report, the better it fits your mail.
How it learns
- Reports. Report spam / Not spam on a message page, Confirm spam / Not spam in the Quarantine, and your users' reports in the user portal. A report trains your organization's corpus and the shared global corpus.
- Auto-learning. Messages that score very high (15 or more, viruses excluded) are learned as spam, and clean messages that score very low (-2 or less) as legitimate mail, in your organization's corpus only.
- Training needs the original message, so reporting only works while a raw copy is kept. Quarantined mail keeps one until it leaves the quarantine; delivered mail only for a short time, so report soon.
- Reporting the same message again with the opposite label flips it: the old counts are removed.
How it scores
Your organization's own counts weigh twice as much as the global corpus, so another customer's newsletters don't look like spam to you. The probabilities become the rules BAYES_00 to BAYES_99; see the rule reference.
The Learning page
Protect > Learning shows one row per corpus: Global (shared) and your organization once it has trained mail.
| Column | Meaning |
|---|---|
| Spam trained / Ham trained | Messages learned as spam and as legitimate mail ("ham"). |
| Tokens | Distinct features stored. |
| Status | learning until both counts reach the minimum (50 by default); then active, and the Bayesian filter starts scoring. |
Getting started faster: report 50 spam and 50 legitimate messages from Message trace early on.