Introspective Users and Introspective Text: Some Recent Results - PowerPoint PPT Presentation

Introspective Users and Introspective Text: Some Recent Results Shomir Wilson – Carnegie Mellon University

Timeline 2 2011: PhD, Computer Science, University of Maryland Metacognition in AI, dialogue systems, detection of mentioned language 2011-2013: Postdoctoral Fellow, Carnegie Mellon University Usable privacy, mobile privacy, regret in online social networks 2013-2014: NSF International Research Fellow, University of Edinburgh 2014-2015: NSF International Research Fellow, Carnegie Mellon University Characterization and detection of metalanguage Also: collaboration with the Usable Privacy Policy Project

Faculty and Professional Collaborators 3 University of Maryland: Don Perlis UMBC: Tim Oates Franklin & Marshall College: Mike Anderson Macquarie University: Robert Dale National University of Singapore: Min-Yen Kan Carnegie Mellon University: Norman Sadeh, Lorrie Cranor, Alessandro Acquisti, Noah Smith, Alan Black University of Edinburgh: Jon Oberlander University of Cambridge: Simone Teufel

Student Collaborators 4 Carnegie Mellon University: Hazim Almuhimedi, Bin Liu, Salem Hilal, Jon Breiger, Rob Murcek, Tommy Doyle University of Cambridge: Kevin Heffernan

Usable Privacy: Motivations 6 http://www.paintsquare.com/blog/images/PSN_1002_Blog_StickyNotes.JPG http://img3.wikia.nocookie.net/__cb20140304030658/degrassi/images/9/9a/ http://stylettomag.co.uk/wp-content/uploads/2014/05/Diary.jpg Big-book.jpg http://thebriberyact.com/wp-content/uploads/2010/09/privacy-policy.jpg

Oversharing, Regret, and Nudging 7 Oversharing in an online social network (OSN) can lead to regret. Can we identify OSN content that individuals are likely to regret? Can we help people maintain their professed sharing preferences? http://time.com/3706434/cella-tweet-fired-texas-jets-pizza/ “I read my Twitter the next morning and was astonished”: A conversational perspective on Twitter regrets. Manya Sleeper, Justin Cranshaw, Patrick Gage Kelley, Blase Ur, Alessandro Acquisti, Lorrie Faith Cranor, Norman Sadeh. CHI 2013.

Twitter Deletion Study 8 OSN post deletion is potentially an indication of regret. Can we study regret via deletion? We tracked 292K active Twitter users for one week and collected their public tweets. We used deletion notices from the Twitter API to track when tweets were deleted. Hazim Almuhimedi, Shomir Wilson, Bin Liu, Norman Sadeh, and Alessandro Acquisti. Tweets are forever: A large-scale quantitative analysis of deleted tweets. In Proc. CSCW 2013.

How Are Deleted Tweets Different? 9 We collected a total of 6.7M tweets. 2.4% were deleted during the observation period. In aggregate, there were some significant differences between deleted and undeleted tweets. Tweet Location: Non-Deletion vs. Deletion Tweet Origin: Non-Deletion (Top) vs. Deletion (Bottom)

Discussion 10 ¨ Deleting a tweet doesn’t mean it’s completely gone ¨ In aggregate, deleted tweets show some intuitive traits ¨ Still, in aggregate, deleted tweets are just barely distinctive

In the Pipeline: A User Study 11 Ideal Scenario: Non-Retweets Reasons for Tweet Deletion Action % Make changes 38 Post nothing 34 No change 23 Other 5 Ideal Scenario: Retweets Action % Do not retweet 47 No change 37 Add comments 13 Other 3

Location Sharing 12

Locaccino (2010-2013) 13 Location sharing and CMU shuttle tracking Available for iPhone and Android ~35,000 downloads Requestor Requestee Time of identity location request Location sharing rule Shomir Wilson, Justin Cranshaw, Norman Sadeh, Alessandro Acquisti, Lorrie Cranor, Jay Springfield, Sae Young Jeong, and Arun Balasubramanian. Privacy manipulation and acclimation in a location sharing application. In Proc. Ubicomp 2013.

Study Motivation 14 Finely-configurable OSN privacy settings are ¤ good: they can reflect users’ nuanced preferences ¤ bad: they require attention to configure and maintain Privacy profiles can represent users preferences. ¤ Mugan et al. clustered OSN users’ location sharing preferences. How does presenting privacy profiles to users influence their comfort with location sharing? Mugan, J., Sharman, T., and Sadeh, N. Understandable Learning of Privacy Preferences Through Default Personas and Suggestions. Technical report CMU-ISR-11-112: Carnegie Mellon University, 2011. Available at http://reports-archive.adm.cs.cmu.edu/anon/isr2011/CMU-ISR-11-112.pdf.

Conditions and Protocol 15 Subjects were randomly assigned to two conditions: ¨ Treatment (“profile”): 16 subjects ¨ Control (“rule”) condition: 18 subjects After initializing their settings, subjects used Locaccino for three weeks. Every night they audited real and hypothetical location sharing requests.

Auditing: Composition of Results 16 Control (“Rule”) Treatment (“Profile”) Request denied, 1 unsatisfied 100% ¡ 100% ¡ Mean ¡Percentgae ¡of ¡Audit ¡Responses ¡ Mean ¡Percentage ¡of ¡Audit ¡Responses ¡ 1 90% ¡ Request allowed, 90% ¡ 1 2 unsatisfied 2 80% ¡ 80% ¡ Request denied, 70% ¡ 70% ¡ 2 3 satisfied 60% ¡ 60% ¡ 3 3 Request allowed, 50% ¡ 50% ¡ 4 satisfied 40% ¡ 40% ¡ 30% ¡ 30% ¡ 4 20% ¡ 20% ¡ 4 10% ¡ 10% ¡ 0% ¡ 0% ¡ Week ¡1 ¡ Week ¡2 ¡ Week ¡3 ¡ Week ¡1 ¡ Week ¡2 ¡ Week ¡3 ¡ Study ¡Week ¡ Study ¡Week ¡

Auditing: Satisfaction Rate 17 Control (“Rule”) Treatment (“Profile”) 100% ¡ 100% ¡ The treatment group Mean ¡Percentgae ¡of ¡Audit ¡Responses ¡ Mean ¡Percentage ¡of ¡Audit ¡Responses ¡ 90% ¡ 90% ¡ experienced a significant 80% ¡ 80% ¡ (p=0.05) increase in 70% ¡ 70% ¡ satisfaction from Week 1 60% ¡ 60% ¡ to Week 3, but the rule condition did not 50% ¡ 50% ¡ (p=0.23). 40% ¡ 40% ¡ 30% ¡ 30% ¡ 20% ¡ 20% ¡ By-week differences between the groups were 10% ¡ 10% ¡ not statistically significant. 0% ¡ 0% ¡ Week ¡1 ¡ Week ¡2 ¡ Week ¡3 ¡ Week ¡1 ¡ Week ¡2 ¡ Week ¡3 ¡ Study ¡Week ¡ Study ¡Week ¡

Auditing: Sharing Rate 18 Control (“Rule”) Treatment (“Profile”) 100% ¡ 100% ¡ Both groups showed Mean ¡Percentgae ¡of ¡Audit ¡Responses ¡ Mean ¡Percentage ¡of ¡Audit ¡Responses ¡ 90% ¡ 90% ¡ trends towards greater 80% ¡ 80% ¡ sharing. 70% ¡ 70% ¡ 60% ¡ 60% ¡ The treatment group 50% ¡ 50% ¡ shared significantly more 40% ¡ 40% ¡ during Week 2 (p=0.01) 30% ¡ 30% ¡ with mild indications of 20% ¡ 20% ¡ the same for Week 1 (p=0.13) and Week 3 10% ¡ 10% ¡ (p=0.093). 0% ¡ 0% ¡ Week ¡1 ¡ Week ¡2 ¡ Week ¡3 ¡ Week ¡1 ¡ Week ¡2 ¡ Week ¡3 ¡ Study ¡Week ¡ Study ¡Week ¡

Discussion 19 Satisfaction in the conditions was roughly equal by the end of the study, but they never converged on an equal quantity of sharing. Privacy profiles, as well as other efforts to simplify privacy choices, can have a significant impact on the levels of privacy that users select.