Facets, But Which Ones?
Source: https://queryunderstanding.com/facets-but-which-ones-4a25c9cf3a1f Author: Daniel Tunkelang
Summary
Not every attribute should become a facet. This article covers strategies for selecting which facets to surface, when to surface them, and how to avoid the presentation bias problem.
Three Strategies for Facet Selection
1. Supply-Based Surface facets that cover the most items in the result set:
- Always shows “full” facets — no misleadingly sparse options
- Problem: may surface attributes users don’t care about (e.g., “item weight”)
2. Demand-Based Surface facets that users click most often across historical sessions:
- Aligns with user intent and real filtering behavior
- Problem: cold-start for new attribute types; perpetuates existing patterns
3. Curation / Editorial Human editors define the canonical facet set per category:
- High quality for core categories
- Doesn’t scale to long-tail categories
- Can be combined with supply/demand signals
The Presentation Bias Problem
If you always show the same facets, you can’t measure demand for facets you didn’t show. Classic exploration-exploitation problem:
- Occasionally surface less-common facets to measure latent demand
- A/B test facet sets
Dynamic Facets
Query-adaptive facets: different queries in the same category may warrant different facets.
- “Red dress” → color facet less useful (already specified); surface occasion/style instead
- “Dress” → color, size, occasion all relevant
Facet Value Ordering
Within a facet, ordering values matters:
- By count (most items first) — shows what’s available
- By predicted relevance — shows what the user probably wants
- Alphabetical — predictable, scannable
Key Concepts
- Supply-based facets — coverage-driven facet selection
- Demand-based facets — click-driven facet selection
- Presentation bias — can’t observe demand for unexposed facets
- Dynamic facets — query-adaptive facet sets
- Facet value ordering — ranking values within a facet