Shared Deliberation in Facebook Support Groups for

Shared Deliberation in Facebook
Support Groups for Sickle Cell Patients
and Caregivers
Rosta Farzan and Keyang Zheng, School of Computing and Information
Aisha Walker,Vascular Medicine Institute
Charles Jonassaint, Department of Medicine
University of Pittsburgh
Social media: an important resource for seeking
health-related information and social support
People influence each other while socializing offline or online.
But… We do not know how!
We aim to answer: How does interactions on online health
support groups like Facebook influence the participants,
especially their decisions and attitudes towards their health
concerns.
Particularly, we are studying two Facebook groups, Sickle Cell Warrior
and Sickle Cell Unite dedicated to Sickle cell patients and caregivers
Users’ discussions on online health discussion
forums often involve strong sentiment
• This sentiment can change as a result of information exchange with
other members
• The change of sentiment can be an indication of members’ influence
on each other
RQ: how does participation in online discussions and
interaction with others converge or diverge the valence of
the discussion?
Research Methods
To answer our research question:
• Collect posts and comments data using Facebook public API
• Generate a labeled dataset using Mechanical Turk
• Developed a computational model to classify each message to
positive or negative.[1]
• Analyze the sentiment changes within interaction around posts
[1] Qiu, B., Zhao, K., Mitra, P., Wu, D., Caragea, C., Yen, J., … Portier, K. (2011). Get Online Support, Feel Better -- Sentiment Analysis and Dynamics in an Online Cancer Survivor Community.
2011 IEEE Third Int’l Conference on Privacy, Security, Risk and Trust and 2011 IEEE Third Int’l Conference on Social Computing, 274–281. https://doi.org/10.1109/PASSAT/SocialCom.2011.127
Gathering posts and comments from Facebook
• With the permission of both groups’ moderator or owner, we
extracted all posts and comments for a period of 10 months
• From Jun 2015 to April 2016
• Sickle Cell Warriors (moderated public Facebook page)
• Sickle Cell Unite (closed private group)
A typical example of a Facebook post interaction
Post and comments come from Sickle Cell Warriors, Inc
Computational model for automatic sentiment
classification of posts and comments
Features of the message:
• Length of the message (word count)
• Average length of words
• Occurrences of people’s name
• Punctuations: question or exclamation
marks
• Positive and negative words
• Strength of the positive & negative
words (e.g. “so supportive”) [2]
Hi warriors, thank you all for being so
supportive and sharing. I got a question
about evenflo. do you use evenflo ? does it
work and how, what exactly it does ? I don't
have pain crisis, but I easily get pneumonia.
Does evenflo work for pneumonia ?
[2] Thelwall, M., Buckley, K., Paltoglou, G. Cai, D., & Kappas, A. (2010). Sentiment strength detection in short informal text. Journal of the American Society for Information Science and
Technology, 61(12), 2544–2558.
Collecting Training Data through Mechanical Turk
• Labeled data is needed for learning what features are associated with
positive or negative sentiment
• Mechanical Turk is a crowdsourcing platform allowing to post micro
task to be completed by a large number of users
• We used Mechanical Turk to label 262 randomly selected Facebook
posts or comments.
• Every posts and comments were judged by 5 different users.
Collecting Training Data through Mechanical Turk
How positive or negative discussion are in
general?
Comments of positive posts
Unite
Unite
Warriors
Total number of posts
2854
1,063
Positive posts
61.3% 52.6%
Negative posts
27.5% 36.7%
Posts with comments
44.3% 58.0%
Warriors
Positive comments
81.1% 71.0%
Negative comments
13.6% 20.7%
Comments of negative posts
Unite
Warriors
Positive comments
69.3% 58.7%
Negative comments
23.9% 33.0%
Sentiment changes over time
All comments in response to a
post are chronologically listed (first
comment would be comment 1)
Replies to comments also
considered as comments in
chronological order they
have been posted
Original
post
1
3
comments
2
4
5
Classify valence of posts and comments
Observe the dynamic of discussions over
time: how sentiment changes as a result of
members’ interaction
Distinguishing dynamics of interaction in
response to positive vs. negative original
post
Applied to posts from each group
separately to compare the dynamics of
public vs. private groups
1
3
2
4
5
Sentiment Probability
Discussions are more likely to converge to
positivity
nth comment
Sentiment Probability
Discussions are more likely to converge to
positivity
nth comment
Sentiment Probability
Discussions are more likely to converge to
positivity
nth comment
Sentiment Probability
Discussions are more likely to converge to
positivity
nth comment
Next step…
• Finer grained analysis
• Specifically, with respect to posts focusing on discussions of therapy
and medications (e.g. HU)
• Tracking repeated participation of the same member
• Key role of participants shifting the valence in one direction of other
Sentiment analysis can be used as a lens to
understand online health support groups
Thank you!
Questions and Comments Please 
Keyang Zheng
[email protected]