This document discusses research on detecting distracted driving content posted on social media platforms. It presents several research questions around understanding the extent of distracted driving posts, how it varies by user demographics, location, and time. It describes collecting a dataset of Snapchat posts and developing classifiers to identify distracted driving images and videos with over 90% accuracy. Key findings include some demographics and times of day having higher rates of distracted driving posts. The research aims to help inform interventions to reduce distracted driving.
Selfie Deaths (KillFies) and Distracted Driving: Social Computing, Machine Learning, and Online Social Media
1. Selfie Deaths (KillFies) and Distracted Driving:
Social Computing, Machine Learning, and Online
Social Media
Cadence, Noida | Nov 19, 2019
Ponnurangam Kumaraguru (“PK”)
Professor of Computer Science, IIIT Delhi
TEDx & ACM Distinguished Speaker
Associate Dean of Student Affairs
linkedin/in/ponguru @ponguru fb.com/ponnurangam.kumaragurupkatiiitd
4. Growth of Selfies
◆ 55% of millennials (18 - 36-year-olds) have posted a "selfie"
on a social media service
◆ 24 billion selfies were uploaded to Google Photos in 2015
alone
4
22. is any non-driving activity that the driver
engages in, which can lead to
1. visual distractions (e.g. taking eyes off the
road),
2. manual distractions (e.g. taking hands off
the driving wheel) or
3. cognitive distractions (e.g. taking the mind
off driving)
23. Distracted Driving Posts on Social Media
391K injuries were reported due to distracted
driving in US [NHTSA, 2016]
14% of all fatal crashes were linked to the use
of smartphones
42% of high school students have texted
while driving
Teenagers and young adults comprised 36%
of casualties.
24. • Individuals spend 30% of their weekly online time on social media.
[GlobalWebIndex, 2018]
• 78% of Internet traffic comes from cell phones. [Nielsen, 2017]
Role of Social Media
25. Sociological Grounding
Goffmann,
1959
“… activity of an individual
which occurs during a period
marked by his continuous
presence before a particular
set of observers and which has
some influence on the
observers”
that many aspects of
Goffman’s approach (e.g.,
impression management)
can work in a framework
that is more aligned to
these spaces, namely
through the metaphor of an
exhibition rather than one
of a stage play.
Hogan,
2010
29. Research Questions
What is the extent of distracted driving content posting
behavior?
Which user demographics correlate with posting distracted
driving content?
31. Research Questions
What is the extent of distracted driving content posting
behavior?
Which user demographics correlate with posting distracted
driving content?
How does distracted driving content posting behavior vary
across cities worldwide?
33. Research Questions
What is the extent of distracted driving content posting
behavior?
Which user demographics correlate with posting distracted
driving content?
How does distracted driving content posting behavior vary
across cities worldwide?
How does distracted driving content posting behavior vary
with time?
34. Research Questions
What is the extent of distracted driving content posting
behavior?
Which user demographics correlate with posting distracted
driving content?
How does distracted driving content posting behavior vary
across cities worldwide?
How does distracted driving content posting behavior vary
with time?
How can we detect distracted driving content
posted on online social platforms ?
35. Data
# Snaps : 6 Million
# Cities: 173
Duration: 1 Month
36. Detecting Distracted Driving
Content
• Manually Annotated Training Set
• 10K snaps (7K negative, 3K positive)
• Fleiss Kappa Agreement Rate: 0.85
• Held out set: 5K snaps
• Two types of classifiers
• Image-Based (sample a frame every second) followed by voting
• Video-Based
42. Contributions
First study to corroborate edgework on online
social platforms.
Proposed initial classifier to work on detecting
distracted driving content. Achieved 92% accuracy.
First study to collect public data available on
Snapchat.
51. 51
Influence of the campaign
◆ Collected tweets with #MainBhiChowkidar
◆ 404,006 tweets posted by 130,290 handles
◆ 46,527 (35.7%) handles added Chowkidar or its
variants to their name
◆ 99 variants of Chowkidar used in names