a selfie is worth a thousand words: mining personal patterns behind user selfie-posting behaviours
TRANSCRIPT
A Selfie is Worth a Thousand Words Mining Personal Patterns behind User Selfie‐posting Behaviours
Tianlang Chen, Yuxiao Chen, Jiebo Luo
University of Rochester
1) Posting selfies on social mediaplatforms is extremely popular inthe social media era2) As a typical online socialbehaviour, people's selfie‐posting behaviour is influencedby people's interest, personality,preference, activity and tendency,and it can in turn reflect andpredict these personal patterns
Our goal in this paper:Exploring the interrelation between the selfie‐posting behaviour and other personal patterns
Framework of this paper:1) Unsupervised WeChat Image Classification 2) User Characterization3) Selfie‐posting Behaviour (SPB) Prediction By User Feature4) Personal Patterns Extraction and Visualization5) Correlation Between High‐level Person Attributes and SPB6) High‐level Person Attribute Prediction By SPB
Overall Framework
Unsupervised Image Classification
Category names and numbers Using Silhouette Coefficient to determine a proper cluster number of K‐means
Evaluation accuracy is 88.5%
User Charactorization
• User‘s frequency feature (F) – A 46‐dimensional vector that reflects the frequency of a user to post different categories other than selfie
• User‘s inertia feature (I)– Reflect user's tendency to upload images of a given category in a Moment
• User‘s singleness feature (S)– Reflect user's tendency to upload images of a single category in a given Moment
Unsupervised Image ClassificationNames/numbers of selfie subcategories
Performance
Four advanced selfie tendency definition
• Group selfie tendency
• Outdoor selfie tendency
• Holding an object selfie tendency
• Face mask selfie tendency
Selfie‐posting Behaviour Prediction
Basic selfie‐posting behaviour prediction
Advanced selfie tendency prediction
Using Text information
Personal Patterns Extraction and VisualizationComparison between selfie‐posting addictand selfie‐posting nonaddict
Personal Patterns Extraction and VisualizationComparison between “a series of selfie”lover and “just one selfie” lover
Personal Patterns Extraction and VisualizationComparison between “It’s all selfies’ loverand “selfie and more” lover
Personal Patterns Extraction and VisualizationComparison of different sets of users withspecial selfie tendencies
Correlation Between High‐level Personal Attributes and SPB
• High‐level Person Attributes Extraction– We cluster users into five types by Non‐negative Matrix Factorization
(NMF) – For each user, we extract six high‐level attributes based on the
clustering result
Six attributes and the categories included
Correlation Between High‐level Person Attributes and SPB
• User Type Description based on High‐level Person Attributes– For a user, the value of an attribute is defined as the sum of frequency
of all categories this attribute includes– For a user type, the value of an attribute is defined as the mean
attribute value of all users this user type includes
Radar plots of attributes of five user types clustered by NMF
Correlation Between High‐level Person Attributes and SPB
Mean values of seven selfie‐posting measures among five user types
Ranking of high‐level attributes for selfie‐posting behaviours
High‐level Person Attribute PredictionUsing user selfie‐posting behaviours feature to predict their high‐level attributes
selfie‐posting behaviour has the potential to indicate whether a user has preference toward certain attributes