Summary of "Week 8.1: Profile Linking on Online Social Media"

This lecture focuses on the problem of profile linking across different online social media platforms, aiming to determine if multiple profiles or handles belong to the same individual. The discussion explores methodologies, challenges, and motivations behind linking user profiles on platforms like Facebook, Twitter, LinkedIn, Instagram, Tumblr, and others.

Main Ideas and Concepts

Problem Statement: Given multiple social media profiles (e.g., Facebook, Twitter, LinkedIn) with different usernames/handles, can we determine if they belong to the same person?
Motivation for Profile Linking:
- Avoid duplicating efforts such as sending advertisements multiple times to the same user across platforms.
- Accurately measure unique user sentiments or opinions across platforms (e.g., political sentiment analysis).
- Aid law enforcement in tracking malicious activities or misinformation spreaders who use multiple accounts.
- Organizations can save resources and better understand their audience or user base.
Challenges in Profile Linking:
- Users often change usernames/handles over time.
- Different platforms provide different types and levels of personal information (e.g., LinkedIn is professional; Tinder is dating-focused; YouTube is content/opinion-focused).
- Some profiles may lack key linking attributes like profile pictures or descriptions.
- Users may deliberately keep accounts independent or anonymous.
- Common names lead to multiple similar handles, complicating linkage.
Data and Observations:
- Large-scale tracking shows 5-7% of users change usernames periodically.
- Profile pictures are frequently changed (20-40% change multiple times within two months).
- Other attributes like descriptions, names, and locations also change but time zones and languages rarely do.
- Usernames are public and relatively easy to track due to unique user IDs assigned by platforms.

Methodologies and Approaches for Profile Linking

Attribute Comparison: Compare common profile attributes across platforms, such as:
- Profile pictures
- Username/handle
- Personal descriptions (e.g., job title, university)
- Friends/followers/followings (social graph)
- Personal websites linked in profiles
Content-Based Analysis:
- Compare posts or shared content across platforms (e.g., same photos posted on Facebook and linked from Twitter).
- Temporal correlation of posts (similar content posted simultaneously).
Historical Data Usage:
- Use past usernames or handles, not just current ones, to track changes over time.
- Track username evolution and reuse patterns.
Graph-Based Techniques:
- Analyze social network graphs (friends, followers) to find overlapping connections.
Self-Identification Links:
- Users sometimes link their own accounts explicitly (e.g., Twitter profile linking to Tumblr or Facebook).
Edit Distance and Similarity Metrics:
- Use string similarity measures (e.g., Jaro distance) to compare usernames that are similar but not identical.
Machine Learning Classifiers:
- Use features extracted from usernames and profile data to train classifiers that predict whether two profiles belong to the same person.
- Example: Using 26 features, a classifier achieved around 76% accuracy.

Practical Insights and Examples

Users often change usernames multiple times, making static matching ineffective.
Profile pictures change frequently and can be a strong but not definitive linking feature.
Explicit self-linking (cross-posting URLs or linking accounts) is a strong indicator of identity.
Some users maintain completely independent accounts, making linkage difficult or impossible with public data alone.
Platforms assign unique user IDs that remain constant despite username changes, enabling tracking over time.

Suggested Activity for Learners

Take two of your own social media accounts (e.g., Facebook and Twitter).
List all features you can use to link these accounts (e.g., profile pictures, descriptions, friends, posts).
List changes you could make to make these accounts appear as the same person.
List changes you could make to make these accounts appear as different people.
Share findings in the course forum.

Speakers/Sources Featured

Primary Speaker: Course instructor (name not explicitly given, referred to as "I" in the lecture).
Researcher Mentioned: Pary (PhD graduate who worked on profile linking and username evolution).
Research Paper Referenced: "Other Times Other Values: Leveraging Attribute History to Link User Profiles Across Online Social Networks."

Summary

This lecture introduces the problem of profile linking across social media platforms, explaining why it is important and how it can be approached using a combination of attribute comparison, content analysis, historical data, graph-based methods, and machine learning. It highlights the dynamic nature of user profiles and the