KnowMe and ShareMe: understanding automatically discovered personality traits from social media and...
Transcript of KnowMe and ShareMe: understanding automatically discovered personality traits from social media and...
Liang Gou, Michelle Zhou & Huahai Yang IBM Almaden Research Center
KnowMe& ShareMe: Understanding Automatically Discovered Personality Traits and User Sharing Preferences
Overview
Background and Questions
Study Method
Results of Validation & Privacy Preferences
Discussion & Conclusion
Personality influences behaviors: occupational proficiency (Barrick & Mount’91) and economic decisions (Ford ’05)
Personality & behaviors
Deriving Personality
“I love food, .., with … together we … in… very…happy.”
Word category: Inclusive Agreeableness
Psycholinguistic studies: personality from text (Yarkoni '10; Tausczik & Pennebaker '10 )
Social Media To PersonalityHundreds of millions of people leave text footprints on social media
Psycholinguistic Analytics
Personality Portrait
This offers opportunities to understand individuals at scale.
Big 5Needs
Values
Two Questions1 How good are the system-derived personality traits?
Derived Traits vs. Users’ Perception
Derived Traits vs. Psychometric Tests
Two Questions(cont.)How would users like to share the derived personality traits in an enterprise context?
What and With whom
Friends Colleagues
Mgr.
Benefits and Risks of Sharing
2
Effect of the User’s Traits
Our Method
The MethodThe Experimental System: KnowMe
Two-part study
Model Validation Sharing Preferences
Big 5 Personality (Golbeck et. al. '11; Yarkoni ’10)
KnowMe
Fundamental Needs (Ford. '05; Yang et. al. '13)
Basic Values (Chen et. al. ’13)
The Survey
Part1: Model Validation• Three sets of psychometric tests
• 50-item Big 5 (IPIP), 26-item basic values (Schwartz ’06), and 52-item fundamental needs (our own)
• Rate the matches with their perception of themselves
The Survey (Cont.)Part2: Sharing PreferencesFor each type of traits, we asked users’ sharing preferences • For four groups
• “public”, “distant colleagues”, “management”, and “close colleagues”
• At three levels • “none”, “range”, and “numeric”
• State the expected benefits and risks in the work place
• and desired controls for sharing their traits
The ParticipantsInvited 1325 colleagues with Twitter presence and also producing at least 200 tweets.
256 completed the study among 625 responses
Source: www.acuteaday.com/blog/category/guinea-pig/
Source: www.backyardchickencoops.com.au/author/kassandra/page/5/
United States (42.0%), Europe (32.1%), other parts of the world (25.9%)
Results
Derived Portrait vs. Psycho-Metric Scores
Correlational analysis of each trait profile (RV Coef: considering all dimensions together within each type of trait)
Over 80% of population, the correlation is statistically significant. • Big 5: 80.8% • Needs: 86.6% • Basic values: 98.21%
Derived Traits vs. User PerceptionAll ratings are above 3 (“moderately matched”) out of 5-likert scale
Overall ratings • Big 5: u=3.4, sd = 1.14 • Values: u= 3.13, sd = 1.17 • Needs: u= 3.39, sd = 1.34
Privacy Preferences: Effects of Traits
Effects of Trait Type • PD(Values) < PD(Big5) or PD(Needs) *** • PD(Neuroticism) < others within Big5 ***
PD(∗) is probability of information disclosure. ◦ p<0.1, * p<0.05, ** p<0.01, *** p<0.001
Effects of Trait Value: tend to share “good things” • PD(High) < PD(Low) ** for traits with “positive names”
• Big 5: Openness (+), Conscientiousness (+) Agreeableness (+)
• PD(Low) < PD(High) ** for traits with “negative names” • Big 5: Neuroticism (-) • Values: Conservation (-), Hedonism (-)
Privacy Preferences: Effects of ActorsEffects of Recipient Type • Overall, 61.5% of partici-
pants were willing to disclose
• Sharing differences are significant: PD(distant/public)< PD(close/mgt.)***.
0.0
0.2
0.4
0.6
close.colleague
management
distant.colleague pub
lic
Group
Percentage Setting
None
Range
Numeric
Privacy Preferences: Effects of Actors(2)Effects of the Sender’s Traits • Certain dimensions of the participants’ personality traits
significantly impact their sharing preferences. • For example,
• Extroversion positively impacts one’s sharing preferences for Big 5 and needs, but not for basic values
• Conscientiousness negatively impacts the sharing of all three types of traits
Perceived Risks and Benefits48.89 %
19.94 %11.63 %
6.79 %5.54 %5.4 %
1.11 %0.69 %Personalized IT Services
Workplace LearningNone
Work FitnessTeaming
Self BrandingSelf Awareness
People Underst. & Inter.
0 20 40 60Percentage
Per
ceiv
ed B
enifi
ts
37.55 %16.03 %
15.19 %10.69 %
7.45 %6.89 %
4.78 %1.41 %Reveal Volunerability
Incomplete ImageLost Privacy
NoneInaccurate Analytics
MisconceptionInformation Abuse
Prejudice
0 20 40 60Percentage
Per
ceiv
ed R
isks
Top Benefits • People understanding
and Interaction • Self Awareness • Self Branding
Top Risks • Prejudice • Information Abuse • Misconception • Inaccurate Analysis
Suggested ControlsTop Controls • Controlled Users • Controlled Data • System Trans - Usage/Function
23.09 %18.13 %
14.29 %12.64 %
7.69 %7.14 %
4.95 %
3.3 %3.3 %
2.75 %2.75 %
GroupingSystem Refresh
Anonymity No Sharing
Controlled TimeOpt OutSecurity
System Trans - FuncSystem Trans - Usage
Controlled DataControlled User
0 10 20Percentage
Sug
gest
ed C
ontro
l
ImplicationsSupport of System Transparency• Clearly explain the meaning of each trait and
intended use !!
!!• Prescriptive and clearly states what it is
capable of and its limitations
“It might happen that people could understand something else from the (trait) name… and this should be explained very carefully”
“ability to mark that certain attributes are inaccurate conveying the inability of system to gauge them properly.”
ImplicationsMixed-Initiative Privacy Preserving
• What to share: Control the granularity of personality traits
• Whom to share with: Be alerted or know when someone is accessing their profiles
• When to share. • Where to share: Sharing channels Source: www.hoax-slayer.com/images/
privacy.jpg
ChallengesData Variety and Model Effectiveness• Multiple Data Source: twitters, facebook
• Multiple Projected “Personality”
Cultural and Language Influence
• Western culture vs. Others / English vs. Others
• Modeling / Interpretation / SharingSource: kimbeach.com/wp-content/uploads/2013/12/Fish-Facing-Challenge.jpg
ConclusionThis work demonstrates the potential feasibility of automatically deriving one’s personality traits from social media with various factors impacting the accuracy of models.
Most people are willing to share their derived traits in the workplace, and a number of factors, including who/whom/when/where, and the perceived benefits/risks, significantly influence the users’ sharing preferences.
• Chen, J., Hsieh, G., Mahmud, J., and Nichols, J. Understanding individuals personal values from social media word use. In ACM Proc. CSCW ’2014.
• Ford, J. K. Brands Laid Bare. John Wiley & Sons, 2005. • Schwartz, S. H. Basic human values: Theory, measurement, and applications.
Revue francaise de sociologie, 2006. • Tausczik, Y. R., and Pennebaker, J. W. The psychological meaning of words: LIWC
and computerized text analysis methods. Journal of Language and Social Psychology 29, 1 (2010), 24–54.
• Yang, H., and Li, Y. Identifying user needs from social media. IBM Tech. Report (2013).
• Yarkoni, T. Personality in 100,000 words: A large-scale analysis of personality and word use among bloggers. J. research in personality 44, 3 (2010), 363–373.
References
Backup
Modeling and Deriving One’s Personality
Why model personality • Psychological characteristics
reflecting individual differences • Consistent and enduring • Link to many aspects in one’s life
• Relationship selection • Problem, emotion coping • Brand/product choices • Occupational proficiency • Team performance
What do we model • Big 5 Personality (OCEAN)
[O’Brien ’96, Neuman ’99, Gosling ’03, Wholan’06]
inventive/curious vs.
consistent/cautious sensitiv
e/nervous v
s.
secure/confident
friendly/compassionate
vs. cold/unkind
outgoing/energetic vs. solitary/reserved
effic
ient
/org
anize
d vs
. eas
y-go
ing/
care
less
Modeling One’s Fundamental Needs (Cont.)
Psychometric empirical studies • Large-scale crowdsourcing of needs scores
and text descriptions from over 2000 people on Mechanical Turks
Statistic analysis to correlate • Psychometric scores with textual
descriptions
Predictive model to derive the needs from one’s tweetsAn example: “Ideal” Positively correlated: accomplish, chauffeur, goal, license, special… Negatively correlated: bad, fix, half, minimum, mix, ugly, wrong, obvious, … (Yang, H. et al. , 2013)
Modeling Basic Human ValuesWhy model human values • Values motivate people and guide their actions • Values transcend specific actions and situations
What do we model • 10-dimensional values as measured through
established psycho-metric surveys
[Schwartz 2006] (Chen, J. et al. , 2013)
RV Coef over Subsets of Population
Effects of Trait Value