Anagha Kulkarni Carnegie Mellon University Jaime Teevan, Krysta M. Svore, Susan Dumais Microsoft Research WSDM 2011 Hong Kong, Feb. 9-12, 2011
Anagha KulkarniCarnegie Mellon University
Jaime Teevan, Krysta M. Svore, Susan DumaisMicrosoft Research
WSDM 2011Hong Kong, Feb. 9-12, 2011
Web search is strongly influenced by time
Queries’ popularity changes over time
Document content changes over time
Query intent (what’s relevant to the user) changes over time
Take away:
Changes in query popularity and
document content can signal
changes in query intent.
Changes in
Query
Popularity
Changes in
Document
Content
Changes in
Query
Intent
http://ncaa.com becomes relevant during the basketball tournament but not before or after
Before During After
http://mahalo.com/march-madness is always about March Madness and is relevant before, during, and after
Before, During, After
100 temporal queries, 20 documents per query
Type of change targeted
Δ query popularity Δ relevant results
Ye
ar
20
09
20
10
Queries that
spiked during this time
period last year
(april fools,
tax extension)
Queries with high
click entropy during this
time period last year
(miss usa, easter
baskets)
Queries related
to scheduled
upcoming events
(ipad, crystal bowersox)
Queries anticipated to
be likely to have
breaking news
(earthquake, lady gaga)
March 25 – May 28, 2010
Query
Frequency (daily)
Document
Number of clicks (per query, daily)
Result rank (per query, daily)
Crawl of document content (daily)
Query Intent
Human relevance judgments (weekly)
Changes in
Query
Popularity
Changes in
Document
Content
Changes in
Query
Intent
Methods
Measure query frequency over time
Characterize query popularity curves
Number of spikes / Periodicity / Shape / Trend
Changes in
Query
Popularity
Changes in
Document
Content
Changes in
Query
Intent
0
0,002
0,004
0,006
0,008
0,01
0,012
0,014
3-25 4-01 4-09 4-15 4-22 4-29 5-07 5-13 5-21 5-28
Qu
ery
Po
pu
lari
ty (
no
rmal
ize
d)
Date
Query Popularity - march madness
During tournament
After tournament
march madnessMultiple spikes / not periodic /
castle / downward trend
0
0,02
0,04
0,06
0,08
0,1
0,12
0,14
0,16
0,18
3-25 4-01 4-09 4-15 4-22 4-29 5-07 5-13 5-21 5-28
Qu
ery
Po
pu
lari
ty (
no
rmali
zed
)
Date
Query Popularity - dancing with the stars
dancing with the starsMultiple spikes / periodic / castle / flat
0
0,01
0,02
0,03
0,04
0,05
0,06
0,07
0,08
3-25 4-01 4-09 4-15 4-22 4-29 5-07 5-13 5-21 5-28
Qu
ery
Po
pu
lari
ty (
no
rmali
zed
)
Date
Query Popularity - mlb
Pre-season During season
mlbMultiple spikes / not periodic /
castle / upward trend
0
0,05
0,1
0,15
0,2
0,25
0,3
3-25 4-01 4-09 4-15 4-22 4-29 5-07 5-13 5-21 5-28
Qu
ery
Po
pu
lari
ty (
no
rmali
zed
)
Date
Query Popularity - miley cyrus
New music video
miley cyrus1 spike / not
periodic / wedge /
flat
0
0,01
0,02
0,03
0,04
0,05
0,06
0,07
0,08
0,09
3-25 4-01 4-09 4-15 4-22 4-29 5-07 5-13 5-21 5-28
Qu
ery
Po
pu
lari
ty (
no
rmali
zed
)
Date
Query Popularity - earthquake
earthquake1 spike / not periodic /
right sail/ flat
Baja CA and Mexico quake
0
0,002
0,004
0,006
0,008
0,01
0,012
0,014
3-25 4-01 4-09 4-15 4-22 4-29 5-07 5-13 5-21 5-28
Qu
ery
Po
pu
lari
ty (
no
rmali
zed
)
Date
Query Popularity - tax extension
tax extension1 spike / not periodic
/ left sail / flat
Tax day
Most queries in study
One or more spikes
Not periodic
Wedge
Show temporal trend
Changes in
Query
Popularity
Changes in
Document
Content
Changes in
Query
Intent
Changes in
Query
Popularity
Changes in
Document
Content
Changes in
Query
Intent
1
10
100
1000
3-25 4-01 4-09 4-15 4-22 4-29 5-07 5-13 5-21 5-28
log(
TF)
mar
ch m
adn
ess
on
pag
e
Date
Content Change - march madness
http://www.mahalo.com/march-madness
http://www.ncaa.com/
During tournament After tournament
61% pages show query-related (TF) change
95% pages changed by less than 15% (Dice)
Periodic and multiple-spike queries are more likely to have content-changed documents
Changes in
Query
Popularity
Changes in
Document
Content
Changes in
Query
Intent
Changes in
Query
Popularity
Changes in
Document
Content
Changes in
Query
Intent
0
1
2
3
4
5
3-25 4-01 4-09 4-15 4-22 4-29 5-07 5-13 5-21 5-28
Hu
man
Re
leva
nce
Ju
dgm
en
ts
Date
Human Judgments - march madness
During: scores/schedules After: general info/planning
Click entropy is correlated with
Spikes
1 spike has lower click entropy than Multiple spikes
Shape
Sail (taxes online, pgatour) has significantly lower click entropy than wedge (robert pattinson, kristen stewart)
Trend
Downwards has lower click entropy
than flat (microsoft update)
Changes in
Query
Popularity
Changes in
Document
Content
Changes in
Query
Intent
Changes in query popularity and document content are related to changes in query intent
• Spikes
• Shape
• Trend
• Click
entropy
• Periodicity
• Human
Relevance
Understanding query intent can be used to improve
Captions, Ranking, Results display
Patterns of query intent change
Zoom: march madness – intent changes from general to specific and back to general
Shift: opening day – different intents throughout year
Static: easter ideas – same intent throughout year
Questions?