[Identity Linkage] Content is King: Identifying Users Across Twitter and Quora Without Private Data

TL;DR

Researchers have developed a framework to link user identities across disparate social platforms like Twitter and Quora using only the text users write. By bypassing private attributes like emails or phone numbers, they achieved up to 84% accuracy for active users using optimized TF-IDF representations. Paradoxically, simple statistical models outperformed deep learning giants like BERT in this specific task.

Context & Positioning

In the era of data privacy, Personal Discrete Attributes (PDAs)—the traditional bread and crumbs of identity linkage—are disappearing behind privacy walls. This paper shifts the focus to intrinsic linguistic features. It is not just about what you say, but the unique "textual footprint" you leave behind. This work sits at the intersection of Digital Forensics, Social Computing, and Natural Language Processing, providing the largest publicly available annotated dataset (27k pairs) for this task.

The "Bag of Words" Strike Back: Methodology

The authors explored the transition from simple vectorization to sophisticated Transformers.

1. The BERT Bottleneck

While BERT is the "SOTA" for many NLP tasks, it struggled here with only 22% accuracy. The authors attribute this to BERT's 512-token limit. Aggregating a user's entire history (thousands of tweets) into a single representation requires heavy truncation (Head+Tail), which discards the very "rare" features needed to distinguish one human from 500 others.

2. TF-IDF & The Power of Idiosyncrasy

The real winner was TF-IDF with 1-2 grams. Why? Because it preserves "noise." Common preprocessing steps like stemming or removing stop-words actually hurt performance. Unique misspellings, specific punctuation habits, and niche vocabulary act as a digital fingerprint that TF-IDF captures perfectly, whereas dense embeddings tend to smooth these signals out.

Table 1: Statistics of the curated Twitter-Quora dataset, highlighting the massive 5.7 billion word corpus.

Experiments: Why Active Users are "Trackable"

The study categorized users by "Prolificacy"—their volume of activity. The results were clear: the more you write, the harder it is to hide.

• Low Activity: Accuracy was significantly lower when users had only 10 tweets.
• High Activity: For users with >5000 tweets and 100 Quora answers, accuracy soared to 84%.

Even when these active users were hidden among a crowd of "normal" users (a 1:5 ratio), the system could still pick them out with 73% precision. This suggests that "Digital Exhaust" is cumulative; your uniqueness as a writer emerges over time.

Fig 1: Confusion matrix for various classification methods. Note that while supervised classifiers are robust, the unsupervised matching approach using TF-IDF yielded the highest raw accuracy.

The Temporal Dimension

The authors also hypothesized that users might post on both platforms simultaneously (temporal correlation). However, they found that while yearly correlations exist, weekly or monthly patterns are too noisy. People use Quora and Twitter for different purposes at different times of the day, making content similarity a much stronger signal than timing.

Fig 3: Pearson correlations across different time scales, showing that yearly activity is the most stable link.

Critical Insight & Conclusion

The core takeaway is a cautionary one for privacy: content is a biometric. Even without your name, your vocabulary and syntax across different platforms (even those as different as Twitter's micro-bursts and Quora's long-form explanations) are linked.

Future Outlook: The authors suggest that moving forward, utilizing vector databases like Facebook Faiss will be necessary to scale this from 500 users to the millions. The "failure" of BERT here also opens a research gap: how do we build Transformers that can process a user's entire multi-year history without losing the "rare word" signals that make them unique?