- Since user already seen results above, and they haven’t clicked and continue reading
- What next the results should be?
- Increase diversity to satisfy all users for reduction of global relevance
Diverse Results Example
- word apple can mean a company or a fruit
- web search should cover both of these meanings (aspects)
Query Reformulations for Web Search Diversification
- xQuAD: Exploiting Query Reformulations for Web Search Result Diversification (2010, Uni of Glasgow)
- Diversify results and hope that at least one will satisfy the user
- query is underspecified, we can find more specific query reformulation
- Various aspects are covered by more query reformulations
- Estimate of coverage of aspects and redundancy of documents
How to Query Reformulations Work?
- How relevant the document to the user, given already seen higher results?
- Each document added to results should cover different aspect
- Previous methods: similarity between docs using maximal marginal relevance
- Paper contribution: similarity between sub-queries
Sub-Query Generation
- query reformulations provided by three major Web search engines
- Created probably via query log mining
- related sub-queries + suggested sub-queries
- relative importance generated sub-queries from centralized ranking of documents covering them
Personalized Re-Ranking
- Managing Popularity Bias in Recommender Systems with Personalized Re-ranking (2019, Uni of Colorado Boulder)
- Document → Item, Query → User, Aspect → long-tail vs short-head
- Goal: Relevant but cover both long-tail (rare) and short-head (popular)
- Use “Smooth” xQuAD - maintain some ratio of long tail items
- personalize based on how much user interacted with long-tail vs short-head items (ratio)
- Vaclav’s opinion: why not make item popularity more continuous instead of using 2 categories?
- Adding small diversification can improve NDCG
Coverage and Submodularity
Coverage is a submodularity and diminishing returns problem - read more here
