A search for “ipad” can rank accessories above iPads when accessories repeat the keyword more often. A signal boosting model corrects this kind of mismatch by using earlier interactions with the same query: documents that users repeatedly chose receive a ranking boost. It is especially useful for head queries, whose high search volume produces enough signals to reveal popular results. section 4.2.2 section 8.0
The basic feedback loop has four steps:
- Log searches and subsequent document interactions, retaining the originating query for each interaction. section 4.2.2
- Group interactions by query and document. In the simplest model, each click contributes one point, so the boost equals that pair’s historical click count. section 4.2.2 section 8.1
- Store the aggregated query, document, and boost values in a sidecar collection alongside the searchable documents. section 4.2.2
- At query time, retrieve matching boost records and add their document-specific weights to the search request. Reaggregate new interactions to update future results. section 4.2.2
From clicks to a boosted search
For the query “ipad”, the basic aggregation joins click events to their originating query events, groups by query and document, and counts clicks: SELECT q.target AS query, c.target AS doc, COUNT(c.target) AS boost FROM signals c LEFT JOIN signals q ON c.query_id = q.query_id WHERE c.type = 'click' AND q.type = 'query' GROUP BY q.target, c.target. The resulting rows can be written to a signals_boosting collection. A query-time lookup orders its matches by boost, takes the top ten, and converts them into document boosts such as "885909457588"^966. In the worked search, boosting moves actual iPads above accessories that previously ranked highly on keyword occurrences. section 4.2.2
Raw clicks are not equally reliable evidence of interest: someone may click merely to browse. An e-commerce variant weights clicks, add-to-cart events, and purchases at 1, 10, and 25 respectively, then sums those weighted counts per query–document pair. These are configurable relative weights, not a universal measure of relevance. Signal age also matters differently across stable product catalogs, short-lived job listings, and fast-changing news. section 8.4 section 8.5
Treat popularity as evidence, not an automatic guarantee of relevance. A basic click-count model is vulnerable to data bias and manipulation. Unlike collaborative filtering, which personalizes results, or learning to rank, which learns weights applicable to unseen queries, signal boosting concentrates on observed query–document interactions—particularly those available for popular queries. section 8.1 section 1.3.3
