A simplified dynamic Bayesian network (SDBN) is a probabilistic click model for estimating search-result relevance from user examination and clicks. It reduces position bias by distinguishing results likely seen from results merely rendered. section 11.2.3
SDBN converts session clicks into query-document relevance estimates through a short procedure:
- Find the lowest-ranked clicked result, or the last click, in each search session. section 11.2.3
- Mark every result at or above that position as examined; leave lower results unexamined. section 11.2.3
- Aggregate these query-document judgments across sessions for learning-to-rank training. section 11.2.3 section 12.1.1
Worked search-session example
For the query "dryer," a session whose last click is rank 9 treats ranks 0 through 9 as examined and ranks 10 and 11 as unexamined. Across sessions, a document with 355 clicks and 2,707 examines receives a grade of about 0.131, while one with 133 clicks and 323 examines receives about 0.412.
A reusable implementation accepts session data containing the query, displayed rank, document identifier, and clicked flag. Production training can add a beta prior: prior_weight controls prior strength, while prior_grade supplies a default relevance probability when evidence is sparse. The resulting judgments can feed feature engineering, pairwise training, model storage, and search ranking. SDBN is simpler and less accurate than DBN, and it does not eliminate presentation bias because users can only act on results shown to them. section 12.1.1 section 11.4
