measure attention

Transcription

measure attention
PATTERNS OF LARGE
SCALE ATTENTION
Mor Naaman
W/ Nir Grinberg, Minsu Park
@informor
PATTERNS OF LARGE
SCALE ATTENTION
Mor Naaman
W/ Nir Grinberg, Minsu Park
@informor
CONNECTED EXPERIENCES
LABORATORY
4
5
Veit et al., ICCV 2016
7
8
Open Questions
?
Forms of
Inputs
?
?
?
Cx
Connector
App
?
?
Contexts
for
Sharing
10
11
AOL/CX: FIRST STEPS
12
Attention.
“A wealth of information creates a
poverty of attention and a need to
allocate that attention efficiently
among the overabundance of
information sources that might
consume it.”
— Herbert Simon
14
Capturing Attention
TRENDS IN ATTENTION
competition
opportunities/tools for capture
mined in physical world
20
TRENDS IN ATTENTION
using finer attention models
21
PAYING ATTENTION
TO ATTENTION
22
STUDY 1: HOW DO
PEOPLE READ (PAY
ATTENTION TO) ONLINE
MEDIA STORIES?
23
43%
65%
What are the key factors that
influence reading depth?
Why do some articles perform better
than others?
26
MEASURE ATTENTION: BENEFITS
understand engagement
rank content
fight clickbait
27
MEASURE ATTENTION: BENEFITS
reason about cognition,
language
predict and understand content
success
28
DATASET
2.3M read events from Chartbeat
9 popular news/media sites
For each event:
Page metadata
User ID
Reading depth
29
What does it look like?
Scrolling
Known
Unknown
8000
6000
CQR 504
Depth (pixels)
DnB
780
DiW 603
4000
BjW 559
2000
BFO 323
0
0
60
120
Time (sec)
180
240
300
MEASURING: MAXIMUM DEPTH (PROPORTION)
: 44.3%
B
: 31.3%
C
: 24.3%
0.3
0.2
0.1
A
B
0.0
% readers
0.4
0.5
A
C
A
20
40
60
80
100
32
Device differences?
mobile
tablet
desktop
Prob. Density Function
2
1
0
Financial, N=5064
2
1
0
2
1
0
0%
Sports, N=5916
Magazine, N=4602
25%
50%
% Read
75%
100%
Figure 3: Distribution of read proportion for the
same articles on mobile (red), tablet (green) and
Referral source?
h?
ce
et
non
hs
he
at
ty
1.5
1.0
0.5
0.0
news
search
social
no ref.
Prob. Density Function
rt
4
d
same articles on mobile (red), tablet (green) and
desktop (blue) devices for 3 exemplary sites.
Financial, N=1900
1.5
1.0
0.5
0.0
1.5
1.0
0.5
0.0
0%
Sports, N=2360
Magazine, N=564
25%
50%
% Read
75%
100%
Figure 4: Distribution of read proportion for the
same articles of readers coming from News (red),
search (green), social (tile) and no referrer (purple)
Can we predict?
Can we predict?
- article's average reading depth
- depth of individual read event
MODEL 1 (PRE PUBLICATION)
length
readability
Sentiment
site
author
topic
read
depth
39
SOME CONTRIBUTING FEATURES…
Length ↓
Author’s historical average ↑
Reading score ↓
Topics: Sports ↑
Science ↓
40
TAKEAWAY: READING
DEPTH/ATTENTION IS
(SOMEWHAT) PREDICTABLE
FROM SOURCE AND
CONTEXTUAL CUES
41
FOLLOW-UP QUESTIONS
Better attention model (reading?
skimming? skipping??)
Language and attention
Narrative and attention
Engagement with other media (YouTube)
42
STUDY 2: HOW DO
PEOPLE WATCH (PAY
ATTENTION TO)
ONLINE VIDEO?
43
TWO DIFFERENT DATASETS
Random YouTube videos,
average view duration
Tracking users, view time for
each video they looked at
45
FOR EACH VIDEO…
actual duration
views
# comments
like ratio
language (sentiment) of comments
like/views, comments/views, share/view
…
46
50% viewed
MIXED-EFFECTS LINEAR REGRESSION
Control for category (humor vs. education)
Control for user
48
SOME EARLY RESULTS
- Length matters (obv)
- Positive association with view count
(thank god)
- Positive association with like/view ratio
(good)
- Negative attitudes matter too:
Negative language in comments pos.
correlated
No impact of like/dislike ratio
49
50
TAKEAWAY:
PROPORTION WATCHED
IS ASSOCIATED WITH
CONTEXTUAL CUES
51
PAYING ATTENTION
TO ATTENTION
52
[email protected]
http://mornaaman.com
@informor