Context-Sensitive Query Auto-completion
Naama Kraus and Ziv Bar-Yossef
1. The Goal
3. Context-Sensitive Approach
Predict the user’s query given a very short
input prefix
6. HybridCompletion
Observation: user context hints to her intent
Contextual resources: visited web pages,
previous queries, clicked results, and more.
galaxyss
u
uranus
uranus
uranus facts
uranus moons
uranus rings
uranus atmosphere
uranus pictures
pluto
Contextual resources:
visited web pages, previous queries, user tweets, and more
NearestCompletions
4. NearestCompletion
Suggest
completions
most
similar
Focuses on recent queries context
Suggests
most
similar toqueries
the user’s recent queries
to completions
the user’s
recent
2. MostPopularCompletion
neptune
State-of-the-art
Suggest most popular queries
Context is not always relevant 
combine context-similar completions
with most popular completions
uranus
ups
Problem: users queries are power law
distributed 
long tail of unpopular queries
uranus
ups
7. Results over real user data
An efficient Nearest Neighbors Search using a standard search library
5. Query Similarity Measure
MostPopular is likely to mis-predict when
given a small number of keystrokes
uranus
pluto
u
usps
ups
ups tracking
utube
united airlines
urban dictionary
In proceedings of WWW’11
semantic
decay
weighted tf-idf:
Ziv Bar-Yossef and Naama Kraus,
Context-sensitive query auto-completion,
In WWW, pages 107–116, 2011