Authorship Attribution in High Cardinality Candidate Pools

51 concepts38 papers1 proposal

Imagine trying to recognize a friend's voice in a quiet room with three other people. It is a trivial task. Now imagine trying to pick out that same voice inside a packed stadium of 50,000 screaming fans.

This is the fundamental hurdle facing modern authorship attribution. Historically, stylometry—the statistical analysis of writing style—operated in small, closed worlds. Models were optimized to distinguish between a handful of historical figures (such as identifying the authors of the disputed Federalist Papers). However, real-world digital forensics, open-source code repositories, web forums, and generative AI tracking require identifying an individual source out of thousands or even millions of possibilities.

When candidate pools scale to this magnitude, traditional machine learning approaches suffer a severe performance collapse. Resolving this requires shifting away from standard classification paradigms toward deep metric learning and explicit stylistic disentanglement.

The Scalability Collapse and Crowded Stylistic Spaces

In small-scale authorship attribution experiments, models frequently achieve near-perfect accuracy. However, Luyckx et al. demonstrated that these setups drastically overestimate a model's real-world performance and feature importance. As the pool of potential authors expands, accuracy degrades significantly.

Why does scale break stylometry? The root cause is that the underlying "stylistic space" becomes increasingly crowded. Writing style is bounded by vocabulary, grammar, and syntax. When a model must choose between ten authors, their stylistic profiles are usually distinct enough to separate. But when the candidate pool grows into the thousands, coincidental stylistic overlaps become inevitable, as observed by Azarbonyad et al. and Sage et al..

This crowding effect spans multiple domains:

  • Source Code: In massive open-source software ecosystems, distinguishing between thousands of developers becomes exceptionally difficult because coding conventions, libraries, and formatting standards overlap heavily (Yang et al.).
  • Online Forums: In active social media threads or web forums, thousands of users post short, informal text snippets, causing individual stylistic signatures to blur together (Solorio et al.; Boenninghoff et al.).
  • Generative AI: As thousands of large language models (LLMs) are deployed, their underlying linguistic footprints begin to share commonalities, making high-precision model attribution increasingly difficult (Uchendu; Zahid et al.).

The Topic Confusion Trap

When a machine learning model is trained to identify authors within a small group, it often takes a shortcut: it memorizes what the authors write about (the topic) rather than how they write (the style). Altakrori et al. formalized this vulnerability through the "topic confusion" task, proving that traditional linguistic markers lose their discriminative power when scaled because different authors inevitably write about the same subjects.

To survive high-cardinality environments, attribution systems must explicitly disentangle content from style.

Standard Model:   [Text] ───> [Shared Representation (Content + Style)] ───> Predicts Author (Fails at scale)

Disjoint Model:   [Text] ───┬─> [Content Latent Space (Topics)]
                            └─> [Stylistic Latent Space (Syntax/Fingerprint)] ───> Predicts Author (Scales)

To achieve this separation, Seroussi et al. developed the Disjoint Author-Document Topic Model (DADT). Instead of mixing all textual features together, DADT isolates document-level content topics from author-specific stylistic topics. By stripping away the topical noise, the model preserves the underlying authorial signature even when thousands of candidates are discussing identical subjects.

Breaking the Multi-Class Bottleneck via Metric Learning

Standard authorship attribution models treat the problem as an $N$-way classification task, where $N$ represents the number of candidate authors. The model processes a text and passes it through a final softmax layer to output a probability distribution across all $N$ classes.

This architecture fails at scale for two reasons:

  1. Computational Complexity: If $N$ grows to tens of thousands, the final layer becomes computationally heavy and statistically fragile.
  2. The Closed-World Constraint: If a new author enters the pool, the entire network must be retrained from scratch to accommodate the new class. Chen noted that even advanced large language models face severe performance degradation and context window constraints when forced to evaluate thousands of candidate profiles simultaneously.

To bypass this bottleneck, researchers have reframed the problem from classification to deep metric learning.

Instead of teaching a network to predict a specific author label, the network is trained to map text into a continuous, high-dimensional latent space where distance corresponds to stylistic similarity.

Classification (Rigid):
[Text] ───> [Neural Network] ───> Softmax over 10,000 fixed authors

Metric Learning (Flexible):
[Text A] ───┐
            ├───> [Shared Encoder] ───> Measure Distance in Latent Space ───> Same or Different?
[Text B] ───┘

Saedi et al. pioneered this shift using Siamese Networks. By training two parallel neural networks with shared weights to evaluate whether two text samples were written by the same person, the system learns a universal stylistic similarity metric. This allows for "one-shot" or open-world verification: the system can evaluate an author it has never seen during training simply by checking where their text lands in the latent space.

Building on this concept, Lee et al. introduced DeepStyle, which utilizes a triplet loss objective. Triplet loss forces the network to pull texts from the same author closer together in the embedding space while pushing texts from distractor authors further away. The authors demonstrated that these embedding-based frameworks are far more resilient to performance decay as the candidate pool grows, allowing systems to successfully navigate thousands of distractors in a crowded latent space. While simple statistical baselines collapse under high cardinality, transformer-based style embeddings maintain their discriminative power across large-scale repositories, as shown by the authors of Cracking Double-Blind Review.

Scaling Synthetic Attribution via Structured Spaces

The challenge of high-cardinality attribution is not confined to human-written text; it mirrors the challenges found in tracking synthetic content. In digital forensics, a critical task is identifying which user or model instance generated a specific AI output.

As An observed, binary detection (determining whether an image or text contains a watermark) is highly robust, but source identification (determining which specific user out of a massive pool generated it) is incredibly fragile. Minor perturbations can cause the identification system to collapse because the unique identifiers begin to collide.

To scale source attribution to hundreds of thousands of users without incurring massive latency or collisions, researchers use structured combinatorial spaces. For example, Li et al. developed a "depth watermark" system for LLMs. Instead of assigning a single static signature to each user, the system partitions the model's vocabulary into sub-lists with varying probability biases. This combinatorial approach allows the system to scale to user pools exceeding 130,000 individuals, maintaining statistical distinctiveness without requiring model retraining.

Whether analyzing the subtle syntactic habits of human writers on a forum or tracking algorithmic watermarks across thousands of AI instances, the core insight of large-scale attribution remains the same: when the candidate pool grows, you cannot rely on simple classification. Success requires isolating invariant stylistic markers and mapping them into structured, continuous metric spaces designed to withstand the pressure of scale.

Key Foundational Work

Authorship Attribution and Verification with Many Authors and Limited Data (2008) — contributed 3 concept elements to this direction.
Learning Pairwise Dissimilarity Profiles for Appearance Recognition in Visual Surveillance (2008) — contributed 1 concept elements to this direction.
System that Identifies Writers (2000) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 38 papers
  2. Version 1 · 2026-10-04 · 37 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_large_scale_authorship_attribution_a2b0fdbf,
  title = {large-scale authorship attribution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/large-scale-authorship-attribution/rdr_a2b0fdbfd35044008119a7b928f177e2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_large_scale_authorship_attribution_a2b0fdbf_v2,
  title = {large-scale authorship attribution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/large-scale-authorship-attribution/rdr_a2b0fdbfd35044008119a7b928f177e2/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake