Analyzing anonymized query and click through logs leads to a better understanding of user behaviors and intentions and provides great opportunities to respond to users with an improved search experience. A large-scale provider of SaaS services, Serials Solutions is uniquely positioned to learn from the dataset of queries aggregated from the Summon service generated by millions of users at hundreds of libraries around the world.
In this session, we will describe our Relevance Metrics Framework and provide examples of insights gained during its development and implementation. We will also cover recent product changes inspired by these insights. Chandra and Susan, from the Summon dev team, will share insights and outcomes from this ongoing process and highlight how analysis of large-scale query logs helps improve the academic research experience.
Actions speak louder than words: Analyzing large-scale query logs to improve the research experience
1. Ac#ons
Speak
Louder
than
Words:
Analyzing
large-‐scale
query
logs
to
improve
the
research
experience
Ted.Diamond
Susan.Price
Raman.Chandrasekar
Code4Lib
2013
2. Overview
• Summon
®
• The
Relevance
Metrics
Framework
(RMF)
• helping
us
go
from
user
acCons
to
metrics
to
a
beEer
user
experience
• Goals
• Data
flow
in
RMF
• Query
Sessions
• From
Logs
to
StaCsCcs
• Metrics
Computed
• Challenges
in
RMF
• Insights
from
RMF
• Summary
2
3. The
Summon®
Discovery
Service
• Hosted/SoQware
as
a
Service
• Match
&
Merge
combines
rich
metadata
and
full
text
from
mulCple
sources
• Single-‐unified
index
• 1
billion+
items
• >
500
customers
• Relevancy
in
>
17
languages
3
5. RMF
Goals
• Observe
and
log
user
ac(ons
• Queries,
Types
of
queries
• Features
used
(e.g.
filters,
advanced
search)
• Click
paEerns
• Compute
quality
of
search
results
• Metrics
from
user
behavior,
such
as
clicks
• Analyze
data
to
improve
search
results
and
enhance
the
research
experience
5
6. Data
Acquired
in
RMF
• Queries
Sample
queries
• Query
terms
• anniversary
9/11
• Filters
• kidney
stone
• moore
vs
mack
• Advanced
syntax
• moore
vs
mack
trucks
• Clicks
•
•
Armee
Deutschland
Rolle
in
Einheit
body
art
in
the
workplace
• the
moplah
rebellion
of
1921
john
j.
banning
• 平凡的世界
We
collect
all
queries
and
• 9780140027631
• “boundary
dispute”
Title:(india)
clicks,
not
just
a
sample
è
• SubjectTerms:"孙少平”
large
logs
• TitleCombined:
(AnalyCcal
Biochemistry)
s.fvf(ContentType,
Book
Review,
t)
• Hawthorne,
Mark.
"The
Tale
of
TaEoos:
The
history
and
culture
of
body
Sampling
will
not
cover
art
in
India
and
abroad.
”
the
long
tail
of
queries
6
7. Key
RMF
concept:
Session
• “Finding”
oQen
spans
mulCple
queries.
• Users
add/remove
filters;
change
search
terms
• We
define
Search
sessions
•
Sequence
of
events
with
same
session
Id,
with:
• No
breaks
>
90
minutes
• Total
elapsed
Cme
<=
8
hours
• Possibly
spanning
day
boundary
• Data
grouped
by
session,
sorted
by
Cme
Different
level
of
abstracCon,
more
robust.
7
8. RMF
Data
Flow
Search
Search
Search
Server
Server
Search
Server
Server
Fetch
Logs
8
9. From
Logs
to
Metrics
&
Sta#s#cs
• Remove
‘noise’
(e.g.
test
queries)
• IdenCfy
session
boundaries
• Associate
clicks
with
queries
• Compute
search
goodness
metrics
for
queries
and
sessions
• Compute
staCsCcs
on
aggregated
data:
Abandonment,
MRR,
DCG
9
11. Metrics
Computed:
Abandonment
Search
Abandonment
l Intui(on:
Good
results
lead
to
clicks
So
compute:
l %
queries
with
no
clicks
on
results
l %
sessions
with
no
clicks
on
results
Usually
lower
abandonment
is
beEer.
11
12. Metrics
Computed:
MRR
Mean
Reciprocal
Rank
(MRR)
• Click
on
result
#3
è
Intui(on:
Relevant
results
MRR
=
0.33
(=
1/3)
should
rank
high
• MRR
=
0.15
è
Compute:
First
good
result
1/(Rank
of
top-‐ranked
clicked
around
rank
6
result)
(~
1/(0.15))
L
Higher
MRR
is
beEer!
• Best
MRR
=
1.0
!
12
13. Metrics
Computed:
DCG
Discounted
CumulaCve
Gain
(DCG)
Intui(on:
Best
to
have
relevant
results
in
the
‘right’
order
So:
l More
points
for
top-‐ranking
results
clicked.
Discounted
as
you
go
down
the
result
set.
l Cumulated
across
all
clicks
for
a
query
l Typical
formula
for
DCG
at
rank
p,
if
ri
is
the
relevance
of
result
at
rank
i:
DCGp
=
r1
+
Σj=2..p
(rj/log2(j))
l We
assume
clicks
imply
relevance
13
14. Challenges
with
Log
Data
• Dealing
with
query
or
click
spam/noise
• Remove
expired
sessions
• Mark
spam
as
“suspect”,
exclude
it
• Note:
If
relaCvely
liEle
spam,
minimal
effect
on
metrics
• Assigning
queries
to
sessions
• Ideally,
a
session
=
one
user
+
one
informaCon
need
• Measuring
relevance
• Clicks:
imperfect
proxy
for
relevance/user
saCsfacCon
• DisCnguishing
real
changes
in
relevance
from
other
causes
(e.g.
academic
calendar)
Be
Pragma(c!
14
16. Impact:
Valuable
Data
Source
Aggregated,
cleaned
data
useful
for
Autocomplete
and
Query
suggesCons
16
17. Impact:
Great
for
Analyses
How
many
results
per
page
(RPP)
is
opCmal?
More
not
always
beEer:
• Too
many
è
user
has
to
wait
• Too
few
è
user
has
to
keep
going
to
the
Next
link
• RPP
was
25,
is
10
OK?
• Used
RMF
to
model
click
rate
changes;
verified
in
producCon
Yes,
users
can
sCll
change
RPP
J
17
18. Session
Abandonment
by
#Terms
in
1st
Query
Abandonment:
Smaller
is
beEer
0
2
4
6
8
10
12
Number
of
search
terms
in
query
18
19. Impact:
Improving
User
Experience
• Abandonment
seen
as
very
different
on
short
vs.
long
queries
• Similar
behavior
seen
in
web
search
• QuesCons:
• Why?
• How
can
we
improve
the
(re)search
experience?
• Web
search:
Need
to
infer
intent,
or
segment
search
Ongoing
work…
19
21. Impact:
Data
Use
Plan
• Huge
variance
in
data
across
Cme
• e.g.
behavior
during,
at
the
end
of,
and
aQer
semester
• [any
guesses
why?]
• Cannot
use
small
segments
of
data
for
decision-‐
making
• Need
to
use
straCfied
samples,
across
Cme
21
22. Takeaways
• Relevance
Metrics
Framework
• Using
what
people
do,
not
say
they
do:
AcCons
vs.
Words
• Sessions
as
a
concept
• Going
from
logs
to
metrics
to
staCsCcs
• Challenges
in
using
this
data
• Some
insights
we
gained:
• Valuable
in
many
ways
è
ConCnual
Improvements
in
Summon
22
23. Thanks!
Thanks
to
Ted,
and
to
the
Summon
team!
Ques#ons??
Contact
us:
Susan.Price@SerialsSolutions.com
Raman.Chandrasekar@SerialsSolutions.com @synthesiser
23