Awesome Search KG

Results Boosting

3 min read

Results Boosting

Definition

Results boosting is a technique for adjusting the relevance scores of search results at query time to promote or demote specific items based on business signals, behavioral data, or business rules — without changing the underlying relevance model.

Boosting operates by multiplying or adding to a document’s baseline relevance score. A boosted document rises in the ranking; a “buried” document falls.

Types of Boosting

Field Value Boosting

Promote documents based on a numeric field:

{
  "function_score": {
    "field_value_factor": {
      "field": "popularity_score",
      "modifier": "log1p",
      "factor": 1.5
    }
  }
}

Common signals:

  • Popularity: sales rank, click count, conversion rate
  • Margin: profit per item
  • Freshness: recency of publication or inventory update
  • Rating: average customer review score
  • Stock level: available-to-ship quantity

Categorical / Rule-Based Boosting

Boost or demote entire categories of results:

  • Promote sponsored products
  • Promote items with free shipping
  • Demote out-of-stock items
  • Exclude discontinued products

Personalized Boosting

Adjust scores per user based on:

  • Purchase history and brand affinity
  • Browsing behavior
  • Loyalty tier
  • Location and delivery constraints

Pinning and Blocking

Extreme forms of boosting:

  • Pin: force a document to a specific rank position (rank 1, 2, …)
  • Block: prevent a document from appearing entirely
  • Bury: push to bottom of results

Boosting vs. Merchandising

Boosting adjusts scores — the ranking algorithm still determines order, but scores are modified. Results Merchandising is a broader practice that includes pinning, campaigns, and manual curation that overrides the ranking entirely.

Boosting in Elasticsearch

Elasticsearch function_score query supports multiple boost functions:

  • field_value_factor — multiply score by a field value
  • gauss/linear/exp decay — proximity-based boost (e.g., distance, recency)
  • script_score — arbitrary scoring logic
  • random_score — controlled randomization for exploration

Multiple functions can be combined with score_mode (multiply, sum, avg, max, min).

Calibration

Boost factors must be calibrated so one signal doesn’t overwhelm relevance:

  • Use logarithmic scaling (log1p) to compress large value ranges
  • Normalize signals to comparable ranges before combining
  • A/B test boost changes to measure revenue impact vs. relevance degradation
  • Monitor for boost gaming: sellers artificially inflating popularity signals

Tradeoffs

BenefitRisk
Aligns ranking with business objectivesCan overwhelm relevance signal
Responds quickly to business changesHard to maintain as rules multiply
No retraining neededMay create feedback loops (popular items get more clicks → more popular)
Transparent and auditableMargin/popularity boost may conflict with user intent

Graph View

Backlinks